Amazon Practice Questions, Discussions & Exam Topics by our Authors

A company wants to launch a new website feature that predicts home prices based on user-supplied home attributes. The attributes include location, square footage, and number of bedrooms and bathrooms. An ML engineer has trained a regression model by using the Amazon SageMaker AI XGBoost algorithm. The model performs well with training data. However, the model significan...

The key issue here is that the model performs well on training data but poorly on real-world validation data. This is a classic generalization problem, typically caused by the training data not representing the true underlying data distribution (data mismatch) or overfitting to training patterns. Correct Option: A) Create a larger training dataset that includes more real-world data. Retrain the model. Why A is correct The biggest gap between training and validation performance usually comes from data distribution mismatch. Adding more real-world representative data helps the model learn patterns that generalize better. This directly improves validation performance without needing complex tuning. In AWS SageMaker XGBoost, better data quality and representativeness often yields the highest ROI compared to hyperparameter tuning. --- Why the other options are incorrect B) Increase the value of the `num_round` hyperparameter `num_round` controls the number of boosting iterations (tre...

Author: CrimsonViperX · Last updated Aug 16, 2026

An ML engineer is designing an AI-powered traffic management system to adjust traffic lights during predicted congestion. The system must use near real-time inference to generate predictions to help prevent traffic collisions. The system must use a batch processing pipeline to perform historical analysis of the predictions to continuously refine and improve the model. The historical analysis will take several hours to evaluate how well the predictio...

We need two capabilities: 1. Near real-time inference with autoscaling for traffic light decisions (low latency, scalable endpoints). 2. Batch/historical analysis pipeline running for hours to evaluate predictions vs outcomes and improve the model. --- ✅ Correct Options A) SageMaker real-time inference endpoints with autoscaling (ConcurrentInvocationsPerInstance) Why this is correct: Amazon SageMaker real-time endpoints are designed for low-latency inference, exactly what traffic prediction needs. The system requires near real-time inference, which is a direct match. Autoscaling using ConcurrentInvocationsPerInstance is the correct AWS-native metric to handle fluctuating traffic demand. This ensures: High availability during peak traffic hours Automatic scaling without manual intervention Low latency response for safety-critical decisions Key factor: real-time inference + correct autoscaling metric --- C) SageMaker Processing job + EventBridge scheduling for batch historical analysis Why this is correct: Amazon SageMaker Processing is specifically designed for: Batch data processing Large-scale evaluation workloads Long-running analysis (hours if needed) The requirement explicitly says: “historical analysis takes several hours” “batch processing pipeline to refine model” Amazon EventBridge is ideal for: Scheduling periodic batch jobs (e.g., daily analysis runs) This setup cleanly separates: Real-time inference (online path) Offline ...

Author: Sofia · Last updated Aug 16, 2026

A company is developing an ML model to forecast future values based on time series data. The dataset includes historical measurements collected at regular intervals and categorical features. The model needs to predict future values based on pas...

The correct choice is: C) Use the Amazon SageMaker AI DeepAR algorithm with matching `context_length` and `prediction_length` hyperparameters. --- Why C is correct This problem is a time series forecasting task, where the model must learn: Historical patterns (trend, seasonality) Temporal dependencies across time steps Future values prediction based on past sequences Amazon SageMaker AI provides a purpose-built algorithm for this: DeepAR, which is specifically designed for probabilistic time series forecasting. Key factors: DeepAR is a supervised deep learning forecasting model It trains on multiple related time series (useful when patterns repeat across entities) It handles categorical features and temporal context It uses: `context_length` → how much historical data the model looks at `prediction_length` → how far into the future it forecasts When to use DeepAR: Demand forecasting (sales, inventory) Weather or sensor forecasting Financial time series prediction Any multi-series forecasting problem --- Why the other options are incorrect A) XGBoost with `scale_pos_weight` XGBoost is primarily a tabular supervised learning algorithm `scale_pos_weight` is used for class imbalance in classificat...

Author: Aarav · Last updated Aug 16, 2026

A company develops a recommendation model and hosts the model on an Amazon SageMaker AI endpoint. The model uses the SageMaker AI endpoint to perform near real-time inference to deliver personalized product recommendations to customers based on browsing history, purchase records, and in-app user interactions. After a major marketing campaign, the company observes a sharp drop in the model's performance...

The problem is asking for a solution that can proactively monitor, detect, and validate model performance issues, especially after a marketing campaign caused a distribution shift. This strongly points to data drift + input data validation + explainability monitoring in Amazon SageMaker. Key services to focus on Amazon SageMaker Model Monitor → detects: Data drift (feature distribution changes) Data quality issues Constraint violations (baseline comparison) Amazon SageMaker Clarify → provides: Feature attribution (explainability) Bias detection and model interpretability insights --- ✅ Correct Answer: A Why A is correct SageMaker Clarify for feature distribution changes After a marketing campaign, user behavior changes → feature distributions shift. Clarify helps with explainability and detecting changes in feature influence, which supports diagnosing why performance dropped. SageMaker Model Monitor for near real-time input validation This is the primary tool for detecting data drift and input anomalies in production endpoints It continuously compares live traffic against a baseline dataset It can alert when distributions deviate significantly (exactly what happened after the campaign) 👉 Together, this provides: Pre-emptive detection of data drift Ongoing validation of input quality Root-cause insights into model degradation --- ❌ Why other options are incorrect B) CloudWatch d...

Author: Leah · Last updated Aug 16, 2026

A company is developing ML models by using PyTorch and TensorFlow estimators with Amazon SageMaker AI. An ML engineer configures the SageMaker AI estimator and now needs to initiate a training job that u...

In Amazon SageMaker AI, the training job is initiated through the Estimator’s `fit()` method. When you configure a SageMaker Estimator (for PyTorch, TensorFlow, or built-in algorithms), you define parameters such as the training image, instance type, input data channels, and hyperparameters. The actual execution of the training job on SageMaker infrastructure only begins when `fit()` is called. ✅ Correct Option: A) `fit` method --- Why `fit()` is correct The `fit()` method is the trigger point for training in the SageMaker SDK. Key factors: It submits a training job to SageMaker. It takes the training dataset location (e.g., S3 URI) as input. It handles: provisioning compute instances copying training data running the training container (PyTorch/TensorFlow) saving the model artifacts to S3 Example scenario: You use `fit()` when: You have defined a SageMaker Estimator You want to start training a model using data in S3 ...

Author: Carlos Garcia · Last updated Aug 16, 2026

An ML engineer is using AWS Glue to transform proprietary data from a third-party vendor to a format that the ML engineer intends to use with the Amazon SageMaker DeepAR forecasting algorithm. The data includes several similar time series data files that the ML engineer must convert to the approp...

To solve this, we need to satisfy two requirements simultaneously: 1. Correct input format for Amazon SageMaker DeepAR 2. Efficient compression for cost-optimized storage in S3 via AWS Glue --- Key facts (exam-relevant) Amazon SageMaker DeepAR supports time-series training data in: JSON Lines (common default) RecordIO-protobuf (MXNet-native format, still valid in ML pipelines) AWS Glue commonly outputs optimized columnar formats like Parquet Compression choices matter: Snappy → fast, Splittable, widely used in big data/ML pipelines gzip → better compression than Snappy but slower and not splittable XZ → very high compression but extremely slow and rarely used in AWS analytics pipelines RecordIO-protobuf + Snappy is a classic ML pipeline combination for efficient training data preparation. --- Option analysis ✅ A) Use Snappy to convert the files to RecordIO-Protobuf and to compress the files. (Correct) Matches DeepAR-compatible format (RecordIO-protobuf) Uses Snappy compression, which is: Fast Widely supported in AWS Glue / big data pipelines Efficient for ML training workflows Best balance of: Compatibility (DeepAR-ready) Performance (fast compression/decompression) Scalability (handles large time-series datasets well) ✔ This is the most AWS-native ML pipeline solution. --- ❌ B) Use XZ to convert the files to RecordIO-Protobuf and to compress the files XZ provides very high compression ratio, but: Extremely slow compression/decompression Not commonly supported in AWS Glue / SageMaker pipelines Adds unnecessary latency for ML training workflows Even if format is correct, compression choice makes it impractical 🚫 Rejected due t...

Author: NebulaEagle11 · Last updated Aug 16, 2026

A company needs to ingest data from data sources into Amazon SageMaker Data Wrangler. The data sources are Amazon S3, Amazon Redshift, and Snowflake. The ingested data must always be up to date with th...

Key requirement analysis The critical requirement here is: data must always be up to date with the latest changes in Amazon S3, Amazon Redshift, and Snowflake while being ingested into Amazon SageMaker Data Wrangler. So we are looking for: No stale snapshots Query-at-source or metadata-driven access Minimal data duplication Managed and scalable connectivity across multiple sources --- Option A) Use direct connections to import data from the data sources into Data Wrangler Why it is incorrect Direct connections in SageMaker Data Wrangler allow users to connect to sources like Amazon S3, Amazon Redshift, and Snowflake and import data. However: In many cases, data is imported as a snapshot into Data Wrangler It is typically used for interactive exploration and preprocessing It does not guarantee continuous synchronization with source changes Refresh must be manually triggered or re-imported When this is used Quick exploratory data analysis One-time or batch preprocessing workflows Small-to-medium datasets where freshness is not critical --- Option B) Use cataloged connections to import data from the data sources into Data Wrangler ✅ Why this is correct Cataloged connections use the AWS Glue Data Catalog as a central metadata layer and integrate with services like: Amazon S3 (via AWS Glue / Athena tables) Amazon Redshift Snowflake (via connectors and federated queries) Key advantages: Data is not necessarily copied upfront Queries can be executed against the latest data in the source systems Provides a governed, centralized metadata layer Enables federated querying and consistent access patterns Supports up-to-date data retrieval at runtime This aligns directly with the requirement: 👉 “always up to date with the latest changes” When this is used Enterprise data lake architectures Multi-source analytics environments Scen...

Author: Krishna · Last updated Aug 16, 2026

A company stores user clickstream data in an Amazon S3 bucket in AWS Account A. The company needs to use the data to train an ML model in Amazon SageMaker AI in AWS Account B. The training will take 10 days. The company needs to use only private IP addresses in the training. The...

The correct answer is B. --- Why B is correct B) Set up a VPC endpoint for the S3 bucket. Set the SageMaker AI OPT_OUT_TRACKING environment variable to 1 in the training job. This option satisfies both critical requirements: 1. Private IP only communication A S3 VPC endpoint (Gateway or Interface endpoint) allows SageMaker in Account B to access S3 in Account A without traversing the public internet. Traffic stays within the AWS private network backbone, satisfying the “only private IP addresses” requirement. 2. No training metadata shared with AWS Setting `SageMaker AI OPT_OUT_TRACKING=1` ensures SageMaker does not send service usage data / metadata for model training tracking to AWS. This is the correct and supported mechanism for opting out. 3. Cross-account access S3 bucket policy in Account A can allow access from the SageMaker execution role in Account B via the VPC endpoint. --- Why other options are incorrect A) VPC peering + email AWS to opt out of metadata collection VPC peering alone is not sufficient for S3 access in a secure/private manner; S3 is not directly reachable via peering unless additional architecture is used. Opting out of metadata collection via email is not a valid AWS mechanism. AWS requires configuration...

Author: ShadowWolf101 · Last updated Aug 16, 2026

An ML engineer wants to use Amazon SageMaker AI to prepare data for training. During exploratory data analysis, the ML engineer notices that several categorical features are missi...

We are dealing with missing values in categorical features during exploratory data analysis (EDA) in Amazon SageMaker AI. Key concept Categorical data = labels/classes (e.g., color, city, product type) For missing categorical values, the standard imputation strategy is: Mode (most frequent value) ✔️ NOT mean/median ❌ (these apply to numerical data) --- Option analysis A) Use SageMaker Clarify to impute categorical features with the mean value. ❌ Incorrect on two levels: Amazon SageMaker Clarify is used for: bias detection explainability It does not perform data imputation Mean is invalid for categorical variables When Clarify is used: Detecting bias in training data Explaining model predictions --- B) Use SageMaker Clarity to impute categorical features with the mode value. ❌ Incorrect because: “Clarity” is not a valid SageMaker service (likely confusion with Clarify) Even if interpreted as Clarify, it still does not do imputation Why it is a trap option: ...

Author: Rahul · Last updated Aug 16, 2026

A company has an ML model in Amazon SageMaker AI. An ML engineer needs to implement a monitoring solution to automatically detect changes in the input data distribution of model features. ...

The requirement is to automatically detect changes in input data distribution (feature drift) in a SageMaker hosted model with the least operational overhead. Key requirement breakdown Goal: Detect input data distribution changes (data drift / data quality issues) Environment: Amazon SageMaker AI endpoint in production Constraint: Least operational overhead → prefer fully managed, built-in monitoring --- ✅ Correct option: A Why A is correct Option A uses Amazon SageMaker AI Model Monitor (Data Quality Monitoring), which is purpose-built for this exact use case. Key mechanisms: Establishes a baseline from training data statistics (mean, std, percentiles, feature distributions) Continuously compares incoming inference data against this baseline Detects: Data drift Schema violations Distribution shifts in features Automatically emits metrics that can be pushed to Amazon CloudWatch CloudWatch alarms provide automated alerting with minimal setup Why this is lowest operational overhead Fully managed monitoring (no custom code needed) Built-in drift detection logic Automatic metric generation Native integration with CloudWatch alarms No manual analysis required --- ❌ Why other options are wrong B) Model quality baseline + Robust comparison Focuses on model quality, not input data distribution Requires ground truth labels, which are n...

Author: Amira99 · Last updated Aug 16, 2026

A company collects customer data every day. The company stores the data as compressed files in an Amazon S3 bucket that is partitioned by date. Every month, analysts download the data, process the data to check the data quality, and then upload the data to Amazon QuickSight dashboards. An ML engineer needs to implement a solution to automatic...

Key requirement analysis The solution must: Automatically check data quality before data reaches QuickSight Work on monthly batch data stored in Amazon S3 Have least operational overhead Be AWS-native and scalable This strongly points to managed, serverless AWS Glue Data Quality capabilities, not custom code or event-driven per-file processing. --- Option A — AWS Glue Crawler + AWS Glue Data Quality rules ✅ (Correct) Why this works AWS Glue crawler updates the AWS Glue Data Catalog monthly Once cataloged, AWS Glue Data Quality (DQ) rules can validate datasets directly Fully managed, serverless, and integrated Minimal operational effort (no custom code, no orchestration complexity) Key advantages No infrastructure management No custom PySpark or Lambda code Works naturally with S3 partitioned data Designed specifically for data quality validation at scale Best use case Batch datasets in S3 Governance/validation before analytics tools like Amazon QuickSight ETL pipelines where schema discovery + validation are needed --- Option B — Glue crawler + Glue job with PySpark custom validation ❌ Why rejected Requires writing and maintaining PySpark code Higher operational overhead (job development, debugging, scaling issues) More complex than native Glue Data Quality rules Not “least operational overhead” ...

Author: Alexander · Last updated Aug 16, 2026

An ML engineer is setting up an Amazon SageMaker AI pipeline for an ML model. The pipeline must automatically initiate a re-training job if any data drift is detected. ...

To meet the requirement—automatic retraining triggered when data drift is detected in an Amazon SageMaker AI pipeline—the solution must use a purpose-built model monitoring capability integrated with SageMaker and a reliable orchestration trigger mechanism. --- ✅ Correct Option: C) Use Amazon SageMaker Model Monitor to detect data drift. Use an AWS Lambda function to automate the re-training job. Why this is correct Amazon SageMaker Model Monitor is specifically designed for ML data quality and model drift detection (e.g., feature distribution changes, data skew, concept drift). It continuously compares baseline training data vs incoming inference data. It can automatically publish drift metrics to Amazon CloudWatch. AWS Lambda is commonly used to: React to CloudWatch alarms or Model Monitor reports Trigger downstream actions such as starting a SageMaker Pipeline or training job Key reasoning factors: Native integration with SageMaker ecosystem Built-in drift detection (no custom logic needed) Event-driven automation via CloudWatch + Lambda Scales well for production ML workflows When this option is used: You want managed ML observability You need automatic retraining based on drift You want minimal custom infrastructure --- ❌ Why other options are incorrect A) AWS Glue crawler + Glue ETL + triggers AWS Glue is designed for data cataloging and ETL pipelines, not ML-specific drift detection....

Author: Nia · Last updated Aug 16, 2026

A company is developing a new ML model to rank customers in order of their potential to pay back loans. The company needs to use an Amazon SageMaker AI built-in algorit...

The correct choice is: A) XGBoost --- Why A is correct: XGBoost Amazon SageMaker XGBoost is a supervised learning algorithm designed for classification and regression problems, which makes it ideal for ranking customers based on their likelihood of repaying loans. Key reasoning factors: The problem is a predictive scoring/ranking task (credit risk / repayment probability). XGBoost handles: Structured/tabular data (customer financial history, income, credit score, etc.) Supervised learning (requires labeled data like “repaid” vs “defaulted”) High accuracy on classification and ranking problems It is one of the most commonly used SageMaker built-in algorithms for credit risk modeling, fraud detection, and churn prediction. When to use XGBoost: Credit scoring models Loan default prediction Customer churn prediction Any tabular supervised learning problem requiring strong performance --- Why other options are incorrect: B) K-means clustering Unsupervised learning algorithm Groups data into clusters based on...

Author: Zain · Last updated Aug 16, 2026

An ML engineer needs to build a processing pipeline to identify and remove personally identifiable information (PII) from petabytes of unstructured data. The ML engineer will use the processed data to ...

The correct choice is A. ✅ Selected Option A: AWS Glue + Detect PII transform (Apache Spark serverless engine) This option uses AWS Glue (serverless Spark ETL) combined with the Detect PII transform, which is specifically designed for large-scale data processing and PII redaction. Why A is correct Scale requirement (petabytes): AWS Glue runs distributed Spark jobs in a serverless, horizontally scalable environment, making it suitable for petabyte-scale unstructured data pipelines. Built-in PII handling: The Detect PII transform can automatically identify and redact sensitive information (names, SSNs, emails, etc.) during ETL. Pipeline fit for SageMaker: Output can be directly written to S3 and used for Amazon SageMaker AI training datasets. Operational advantage: No cluster management; fully managed ETL job execution. 👉 When to use this: Use AWS Glue when you need large-scale ETL, data cleaning, or transformation pipelines over massive datasets (TB–PB scale) with built-in transformations like PII detection. --- ❌ Why other options are incorrect B) AWS Glue Data Wrangler in Amazon SageMaker Canvas Best for: No-code/low-code data preparation and exploration. Limitation: Designed for small to medium datasets, not petabyte-scale pipelines. Why rejected: Can...

Author: Oliver · Last updated Aug 16, 2026

An airline company uses an ML model to adjust ticket prices based on demand. The model runs on Amazon SageMaker real-time endpoints. During previous deployments, the model failed to scale quickly enough when website traffic increased, which caused delays in price adjustments. An ML engineer needs to configure auto scaling for the SageMaker endpoints to respond ra...

We need the most responsive auto scaling configuration for Amazon SageMaker real-time endpoints using target tracking, optimized for sudden traffic spikes. Key concepts for AWS exam reasoning For SageMaker endpoint auto scaling, responsiveness depends on: 1. Metric choice `InvocationsPerInstance` is the correct target tracking metric for inference workloads. It reflects real-time request load per instance → directly tied to scaling decisions. 2. Metric resolution Standard resolution → 1-minute granularity → slower reaction. High-resolution (10-second) → near real-time signal → faster scaling reaction. 3. Cooldown periods Scale-in cooldown prevents removing capacity too quickly. Lower cooldown → more responsive scaling (faster scale-in adjustments when traffic drops). Higher cooldown → slower reaction, risk of overprovisioning or lagging adjustments. --- Evaluate options ❌ A Standard metric (1-minute granularity) → slow detection of spikes 10-second interval mentioned, but standard metric limits responsiveness 300s cooldown is reasonable, but metric resolution bottleneck dominates ➡️ Rejected because standard metric reduces responsiveness significantly --- ❌ B High-resolution metric (good 👍) BUT 600-second scale-in cooldown → very slow to scale in System becomes sluggish in a...

Author: Suresh · Last updated Aug 16, 2026

A company uses ML models to predict whether transactions are fraudulent. The company needs to identify as many fraudulent transactions as possible. Which evaluation metr...

We first translate the business requirement into a classification objective: The company wants to identify as many fraudulent transactions as possible. This means the primary goal is to minimize false negatives (fraud cases that are missed). In other words, the model should prioritize catching all actual fraud, even if it sometimes incorrectly flags legitimate transactions. Correct Option: Recall Recall is defined as: > Recall = True Positives / (True Positives + False Negatives) It directly measures how many actual fraud cases the model successfully detects. A high recall means very few fraudulent transactions are missed. This makes recall the most appropriate metric when the cost of missing a fraud case is high. --- Why other options are rejected A) F1 Score F1 score is the harmonic mean of precision and recall. It is useful when you need a balance between false positives and false negatives. However, the question does not ask for balance; it explicitly prioritizes catching as many frauds as possible. Therefore, F1 is not optimal because it still penalizes recall improvement if precision drops. Use case: When both false positives and false negatives are equally important, such as general classification ...

Author: CrimsonViperX · Last updated Aug 16, 2026

A company has significantly increased the amount of data that is stored as .csv files in an Amazon S3 bucket. Data transformation scripts and queries are now taking much longer than they used to take. An ML engineer must implement a solution to optimize th...

This is a classic “optimize S3 data for analytics/query performance with least operational overhead” question. The key is not just making files smaller, but choosing a storage format and processing approach that improves scan efficiency. --- Key requirement breakdown Data is currently .csv in Amazon S3 Queries and transformations are slowing down Goal: optimize query performance Constraint: least operational overhead So we want: Reduced data scan time Column pruning support Compression + efficient I/O Minimal cluster management --- ✅ Correct Answer: C) AWS Glue ETL job to convert CSV to Apache Parquet Why C is correct Converting CSV → Apache Parquet is one of the highest-impact optimizations for analytics workloads. Key reasons: Columnar format (Parquet) → only reads required columns (not full row scan like CSV) Compression (Snappy/ZSTD) → smaller storage + faster I/O Predicate pushdown → filters applied at storage level Works seamlessly with Amazon Athena, AWS Glue, Amazon Redshift Spectrum Fully managed serverless ETL using AWS Glue → low operational overhead When to use this: Large datasets in S3 used for analytics Athena/Glue/Redshift Spectrum queries Need long-term scalable optimization of query performance --- ❌ Why other options are wrong A) AWS Lambda to split CSV files into smaller objects Why it seems tempting: Smaller files can improve parallelism in reads Why it is NOT best: Does not change file format Still CSV → inefficient scanning remains Adds custom orchestratio...

Author: Isabella1 · Last updated Aug 16, 2026

An ML model is deployed in production. The model has performed well and has met its metric thresholds for months. An ML engineer who is monitoring the model observes a sudden degradation. The performance metrics of the...

The correct answer is: B) Drift in production data distribution --- Why B is correct (key concept: data drift / concept drift) A model that has been performing well for months and suddenly drops below performance thresholds in production is a classic sign of data drift (also called covariate shift or concept drift depending on what changed). In production ML systems, the most common reason for sudden degradation is: The statistical distribution of incoming real-world data changes over time The model was trained on an older distribution and is now seeing different patterns This leads to reduced accuracy, precision, recall, etc. Example scenario where B applies: A fraud detection model trained on last year’s transaction patterns Suddenly new fraud techniques emerge Input patterns change → model predictions become unreliable This is exactly why AWS services like Amazon SageMaker Model Monitor track drift in features and predictions. --- Why other options are incorrect: A) Lack of training data ❌ If there was insufficient training data, performance issues would appear during training or initial deployment, not after months of stable production. Once deployed and stable, training data quantity d...

Author: CrystalWolfX · Last updated Aug 16, 2026

A hospital is using an ML model to validate x-ray results. The hospital runs a nightly batch inference job. The hospital needs to produce a daily report about model data qua...

The requirement is to produce a daily report on model data quality and model performance for a nightly batch inference job. This points directly to a managed ML monitoring solution rather than general observability or ETL data checks. --- ✅ Correct Option: A A) Schedule a monitoring job in Amazon SageMaker Model Monitor. Generate the monitoring results for the model and data. This is the best fit because: It is purpose-built for ML model monitoring in production Supports: Data quality monitoring (schema violations, feature drift) Model quality monitoring (accuracy, precision/recall when ground truth is available) Bias and feature drift detection Works well with batch inference pipelines, where predictions are logged and analyzed daily Can be scheduled to run automatically (e.g., daily after batch job completion) Produces monitoring reports suitable for dashboards or audit reporting 👉 Key alignment with requirements: Nightly batch inference ✔ Daily reporting ✔ Data + model performance monitoring ✔ --- ❌ Why other options are incorrect B) CloudWatch dashboard for batch inference metrics Using Amazon CloudWatch dashboards only provides: CPU/memory usage job success/failure latency or throughput 🚫 Why it’s wrong: Does not analyze model quality Does not detect data drift or prediction accuracy degradation Only infrastructure/operational monitoring, not ML-specific validation 👉 When this is used: ...

Author: Sam · Last updated Aug 16, 2026

A company runs an Amazon SageMaker AI domain in a public subnet of a newly created VPC. The network is configured properly, and ML engineers can access the SageMaker AI domain. Recently, the company discovered suspicious traffic to the domain from a specific IP address. The company n...

Key requirement: block traffic from a specific IP address at the network layer for a SageMaker AI domain in a VPC public subnet. In AWS networking, the two main controls for IP-level filtering are Security Groups (stateful, allow-only) and Network ACLs (stateless, allow + deny). Route tables and SageMaker inference features do not provide IP-based filtering. --- ✅ Correct option: B B) Create a network ACL inbound rule to deny traffic from the specific IP address. Assign the rule to the default network ACL for the subnet where the domain is located. A Network ACL (NACL) is the only option here that supports an explicit DENY rule for a specific IP address. Since NACLs are associated with subnets, applying it to the subnet hosting the SageMaker domain effectively blocks that IP before traffic reaches the instances. Key factors: NACLs are stateless → evaluate both inbound and outbound traffic separately Support explicit DENY rules Operate at the subnet level, making them suitable for coarse-grained IP blocking Rule order (lowest number first) is important for evaluation ...

Author: Mia · Last updated Aug 16, 2026

A company is using Amazon SageMaker AI to develop a credit risk assessment model. During model validation, the company finds that the model achieves 82% accuracy on the validation data. However, the model achieved 99% accuracy on the training data. The com...

The problem shows a classic overfitting scenario: Training accuracy: 99% (very high) Validation accuracy: 82% (significantly lower) This indicates the model has learned the training data too well, including noise and patterns that do not generalize. --- ✅ Correct Answer: B) Implement dropout layers. Use L1 or L2 regularization. Perform k-fold cross-validation. Why B is correct This option directly targets overfitting (high variance) using proven generalization techniques: Dropout layers Randomly deactivate neurons during training → prevents co-adaptation of features → improves generalization. L1/L2 regularization Adds penalty for large weights → discourages overly complex models → reduces variance. k-fold cross-validation Evaluates model on multiple train/validation splits → ensures performance is stable across datasets → reduces risk of overfitting to a single split. 👉 These are standard AWS ML best practices when training models in Amazon SageMaker for improving generalization. --- ❌ Why other options are incorrect A) Add dense layers + batch normalization + early stopping ❌ Adding more dense layers increases model complexity, which worsens overfitting in this scenario. ✔ Batch normalization helps stabilize training. ✔ Early stopping helps prevent over-training. 👉 But overall, this option is mixed and incorrectly pushes...

Author: Arjun · Last updated Aug 16, 2026

A company has developed a computer vision model. The company needs to deploy the model into production on Amazon SageMaker AI. The company has not hosted a model on SageMaker AI previously. An ML engineer needs to implement a solution to track model versions. The solution also must provide ...

The correct answer is C) Register the model in the SageMaker Model Registry. Use SageMaker Inference Recommender for recommendations about instance types. Why option C is correct This question has two key requirements: 1. Model version tracking 2. Recommendation of optimal EC2 instance types for hosting inference 1. Model version tracking → SageMaker Model Registry The Amazon SageMaker Model Registry is specifically designed for: Tracking multiple versions of models Storing model artifacts, metadata, and lineage Supporting approval workflows before deployment Integrating with SageMaker endpoints for deployment Since this is the company's first time hosting a model on SageMaker AI, the Model Registry is the standard and correct service for managing production-ready model versions. 2. Instance type recommendation → SageMaker Inference Recommender SageMaker Inference Recommender is purpose-built for: Benchmarking model performance across different instance types Running load tests for latency, throughput, and cost optimization Providing data-driven recommendations for best EC2 instance types Reducing guesswork when selecting hosting infrastructure This directly matches the requirement to recommend EC2 instance types for model hosting. --- Why other options are incorrect ❌ A) ECR + Compute Optimizer Amazon ECR stores container images, not model versions. It does not provide model lifecycle tracking. AWS Compute Optimizer recommends EC2 sizin...

Author: Lina Zhang · Last updated Aug 16, 2026

An ML engineer wants to use, prepare, and load data from Amazon S3 for analytics. The ML engineer must run an extract, transform, and load (ETL) job to discover the schema of the data and to store the...

Correct Answer: A) Use AWS Glue to run the ETL job. Use the job to discover the schema and to store the associated metadata in the AWS Glue Data Catalog. --- Why Option A is correct (AWS Glue) This question is centered on least manual effort + automated schema discovery + metadata cataloging for S3 data, which strongly points to AWS Glue. Key factors: Fully managed ETL service → no infrastructure to manage Automatic schema discovery (crawlers) for data in Amazon S3 Built-in Data Catalog to store metadata centrally Designed specifically for ETL workloads over data lakes (S3) Integrates natively with analytics tools like Athena, Redshift, and SageMaker 👉 In this scenario: Glue Crawler scans S3 data Infers schema automatically Stores metadata in AWS Glue Data Catalog Glue ETL job transforms data if needed This is exactly what the question asks: ETL + schema discovery + metadata storage with minimal effort --- Why other options are incorrect ❌ B) SageMaker Data Wrangler + store metadata in S3 Data Wrangler is mainly for: Visual data preparation for ML Prototyping feature engineering Not designed for enterprise ETL pipelines Does NOT provide a centralized metadata catalog like Glue Storing met...

Author: Daniel · Last updated Aug 16, 2026

A company has trained an ML model that is packaged in a container. The company will integrate the model with an existing Python web application. The company needs to host the model on AWS by using Kubernetes. The company does not want to manage the control plane and must provision the resources...

The requirements in this question are driven by three key constraints: 1. Kubernetes without managing the control plane → This strongly points to a managed Kubernetes service such as Amazon Web Services ’s Amazon Elastic Kubernetes Service (EKS). 2. Repeatable infrastructure provisioning → implies Infrastructure as Code (IaC), not manual CLI steps. 3. Provisioning must be done using Python → narrows the solution to a Python-native IaC tool. Also, the model container must be stored in a registry like Amazon Elastic Container Registry (ECR) for deployment into Kubernetes. --- Option Analysis A) CloudFormation + EC2-based Kubernetes cluster This implies a self-managed Kubernetes cluster on EC2. Problem: You must manage the Kubernetes control plane yourself (or simulate it), which violates the requirement: “does not want to manage the control plane.” Also, EC2-based Kubernetes is operationally heavy and error-prone compared to EKS. Even though CloudFormation is repeatable IaC, it is not Python-based. When this is used: Legacy environments requiring full control over Kubernetes components. On-prem-like setups in AWS. ❌ Rejected due to self-managed control plane + not Python-based provisioning. --- B) AWS CLI + EKS + ECR Uses EKS, which is correct for managed Kubernetes. However, AWS CLI is not Infrastructure as Code: Not declarative Not version-controlled in a structured way Not easily repeatable for full environments...

Author: Sara · Last updated Aug 16, 2026

A logistics company has installed in-vehicle cameras for basic monitoring of its drivers. The company wants to improve driver safety by identifying distractions that could lead to accidents...

The requirement is to identify driver distractions from in-vehicle camera footage with the least operational effort. This strongly points toward a fully managed computer vision service that already provides pre-trained models for face and behavior analysis. --- ✅ Correct Option: A) Use Amazon Rekognition eye gaze direction detection Why this is correct Amazon Rekognition is a fully managed AI service that can analyze images and video without requiring you to build or train custom models. Key factors: Low operational effort: No model training or infrastructure management required Pre-built facial analysis: Includes face detection, landmarks, head pose estimation, and can be used to infer attention/gaze direction Real-time video processing: Works directly with camera feeds (e.g., in-vehicle cameras) Scales easily across fleets of vehicles When this option is used Detecting driver attention (looking forward vs away from road) Identifying fatigue or distraction indicators from video Monitoring safety compliance in transport/logistics fleets --- ❌ Why other options are incorrect B) Use Amazon SageMaker AI to customize a model Amazon SageMaker Requires building, training, and maintaining custom ML models Needs data...

Author: Grace · Last updated Aug 16, 2026

An ML engineer uses one ML framework to train multiple ML models. The ML engineer needs to optimize the inference costs and host the models on Amazon SageMaker AI....

The correct answer is B) Create a multi-model inference endpoint for all the models. Why B is correct (Key reasoning) A SageMaker Multi-Model Endpoint (MME) is designed specifically for this scenario: Multiple models are trained using the same ML framework All models share a single serving container Models are loaded dynamically from Amazon S3 as needed Only active models are kept in memory, reducing idle resource cost Why this is MOST cost-effective Single endpoint + shared compute resources Avoids provisioning separate instances per model Scales efficiently when there are many models with variable traffic Reduces infrastructure duplication and idle capacity costs --- Why the other options are incorrect A) Multi-container inference endpoint (direct invocation) Used when you need multiple containers in parallel within one endpoint Example use: ensemble models or different frameworks in one request flow ❌ Not designed for hosting ma...

Author: FlamePhoenix2025 · Last updated Aug 16, 2026

An ML engineer is analyzing a classification dataset before training a model in Amazon SageMarker AI. The ML engineer suspects that the dataset has a significant imbalance between class labels that could lead to biased model predictions. To confirm class imbalance, t...

To detect class imbalance before training in an Amazon SageMaker AI dataset, we need a pre-training bias metric that directly measures how unevenly labels are distributed. --- ✅ Correct Answer: B) Difference in proportions of labels (DPL) Why DPL is correct Difference in Proportions of Labels (DPL) is specifically designed to measure class imbalance. It compares the proportion of each class in the dataset (e.g., positive vs negative labels). If one class is significantly more frequent than another, DPL will reflect a high imbalance value. This is exactly what is needed for pre-training bias detection in classification datasets in SageMaker Clarify. 👉 In AWS SageMaker Clarify, DPL is commonly used as a dataset-level bias metric before training to ensure fairness and balanced representation. --- ❌ Why other options are incorrect A) Mean Squared Error (MSE) Measures prediction error between predicted and actual continuous values. Used in regression problems, not for dataset bias or class imbalance. Does not analyze label distribution at all. 👉 Use case: Model evaluation for regression tasks (e.g., house price prediction). ...

Author: Michael · Last updated Aug 16, 2026

An ML engineer is building an ML model in Amazon SageMaker AI. The ML engineer needs to load historical data directly from Amazon S3, Amazon Athena, and Snowf...

The correct answer is D) Use Amazon SageMaker Data Wrangler to query and import the data. Why D is correct Amazon SageMaker Data Wrangler is specifically designed to connect directly to multiple data sources and import data for ML workflows. It supports native or connector-based access to: Amazon S3 (historical datasets) Amazon Athena (SQL queries over S3 data) Snowflake (via built-in JDBC/connector integration) Key reasoning factors: It provides direct data ingestion + transformation in one tool It is built for ML dataset preparation inside SageMaker It supports multi-source querying without needing separate ETL pipelines It integrates seamlessly with training workflows in SageMaker AI This exactly matches the requirement: load historical data directly from S3, Athena, and Snowflake into SageMaker AI. --- Why other options are incorrect A) AWS Glue DataBrew AWS Glue DataBrew DataBrew is mainly for visual data cleaning and transformation It is typically S3-centric and not designed as a unified ingestion layer for SageMaker training workflows It does not natively provide the same seamless multi-source ML ingestion experience (especially Snowflake + Athena together) When it is used: Data cleaning and preprocessing datasets stored primarily in S3 Business a...

Author: Emma · Last updated Aug 16, 2026

A healthcare company wants to detect irregularities in patient vital signs that could indicate early signs of a medical condition. The company has an unlabeled dataset that includes patient health records, medication history, and life...

We need to identify the best AWS SageMaker algorithm for detecting irregularities (anomalies) in unlabeled patient time-series/health data. Key requirement breakdown Unlabeled dataset → rules out supervised learning (no target labels). Goal: detect irregularities / anomalies in patient vitals → anomaly detection problem. Data includes time-based health records + lifestyle + medication history → likely mixed structured/time-series signals. Need: unsupervised anomaly detection, not prediction or classification. --- Option A: XGBoost (SageMaker XGBoost) Why it is incorrect: XGBoost is a supervised learning algorithm (classification/regression). It requires labeled target variables (normal vs abnormal). Even though it is powerful, it cannot directly detect anomalies in unlabeled data. Hyperparameter issue: `max_depth > 100` is unrealistic and harmful. It leads to overfitting, extremely complex trees, and poor generalization. Typical values are 3–10 in practice. When XGBoost is used: Fraud detection, disease prediction, churn prediction when labels exist. --- Option B: k-means clustering Why it is partially relevant but incorrect for this use case: k-means is unsupervised, so it works with unlabeled data. However, it is designed for clustering, not anomaly detection directly. It groups similar patients but does not inherently identify “irregularities.” You would need additional logic (e.g., distance from centroid thresholding) to detect anomalies. Hyperparameter issue: `k` is required, but selecting k is non-trivial and domain-dependent. Incorrect k can distort clustering and hide anomalies....

Author: Max · Last updated Aug 16, 2026

A company is using Amazon SageMaker AI to deploy a new recommendation model for its ecommerce website. The model must use data from all client website interactions as input. Traffic is variable throughout the day. The company needs to create an inference ...

To choose the most cost-effective SageMaker inference option, we focus on traffic variability, cost model (always-on vs scale-to-zero), and request pattern (real-time vs queued vs batch). ✅ Correct Choice: D) Serverless inference endpoint A SageMaker Serverless Inference endpoint is the best fit because it: Automatically scales from zero to accommodate traffic spikes Charges only for actual compute time used (per request) Eliminates idle costs, which is critical when traffic is highly variable throughout the day Works well for event-driven or intermittent real-time requests, such as ecommerce recommendations triggered by user activity This makes it the most cost-efficient option when traffic is unpredictable but still requires real-time responses. --- Why the other options are rejected A) Batch transform inference endpoint ❌ Designed for offline batch processing, not real-time requests Processes data stored in S3 in bulk jobs Suitable for: Nightly recommendation generation Offline model scoring Not suitable here because the website needs real-time or near-re...

Author: Harper · Last updated Aug 16, 2026

A company runs a neural network model and retrains the model when the performance degrades. The company uses a training job that uses Amazon SageMaker AI distributed data parallelism (DDP). The training job takes several hours to run. The compa...

We are asked how to reduce training time for an Amazon SageMaker AI model that already uses Distributed Data Parallelism (DDP) and currently takes several hours. Key idea With DDP, the main lever for reducing training time is scaling out compute (more instances) so that training data is processed in parallel across more workers. --- ✅ Correct Option: D) Increase the number of instances Why this is correct In SageMaker distributed data parallelism, each instance (worker node) trains on a subset of the data and synchronizes gradients. If you increase the number of instances, you: Split the dataset across more nodes Reduce per-node workload Decrease total training time (up to communication overhead limits) When this is used Large datasets Long training jobs When GPU/CPU utilization per instance is already high When model architecture is fixed but training is slow --- ❌ Why other options are incorrect A) Increase the numb...

Author: Amira99 · Last updated Aug 16, 2026

A company needs to deploy a custom-trained classification ML model on AWS. The model must make near real-time predictions with low latency and must handle vari...

The requirement is near real-time predictions with low latency and the ability to handle variable request volumes, which strongly points toward a managed real-time inference service with autoscaling. Correct option: C) Deploy an Amazon SageMaker AI endpoint. Configure auto scaling for the endpoint. Why C is correct: Low latency inference: Amazon SageMaker real-time endpoints are designed specifically for sub-second, near real-time predictions. Elastic scaling: Auto Scaling adjusts the number of instances based on traffic, handling variable request volumes efficiently. Fully managed: AWS handles provisioning, deployment, monitoring, and scaling infrastructure. Production-ready ML serving: Built for continuous, online inference use cases. This is the standard AWS architecture for production-grade ML inference APIs. --- Why the other options are incorrect A) SageMaker batch transform Not real-time: It processes data in batches, not individual requests. Best use case: Offline scoring Large datasets (e.g., daily predictions, report generation) Fails requirement: cannot provide low-...

Author: Elizabeth · Last updated Aug 16, 2026

An ML engineer is building a logistic regression model to predict customer churn for subscription services. The ML engineer is using a dataset that contains two string variables: location and job_seniority_level. The location variable has 3 distinct values, and the job_seniority_level variable has over...

For a logistic regression model, proper preprocessing of categorical features is critical because the model assumes numeric input without unintended ordinal relationships. We evaluate each option based on the nature of the variables: location: 3 distinct categories (nominal, no inherent order) job_seniority_level: >10 categories (typically ordinal: junior → mid → senior → lead, etc.) --- ✅ Correct Option: B B) Apply one-hot encoding to location. Apply ordinal encoding to job_seniority_level Why this works: 1. Location → One-hot encoding (correct choice) Location is a nominal categorical variable No natural ordering (e.g., Chennai ≠ Mumbai ≠ Delhi in rank) One-hot encoding prevents the model from assuming any hierarchy Works well in logistic regression since it converts categories into independent binary features When to use: Small to moderate cardinality categorical variables Nominal data (no order) Linear models (logistic regression, linear regression) --- 2. Job_seniority_level → Ordinal encoding (correct choice) Seniority levels usually have a meaningful order (e.g., Intern < Junior < Senior < Lead) Ordinal encoding preserves this relationship by mapping to integers When to use: Ordered categorical variables Features where ranking matters --...

Author: Noah · Last updated Aug 16, 2026

An ML engineer is collecting data to train a classification ML model by using Amazon SageMaker AI. The target column can have two possible values: Class A or Class B. The ML engineer wants to ensure that the number of samples for both Class A and Class B are balanced, without losing any exi...

The requirement has two parts: 1. Test whether the dataset is balanced (Class A vs Class B) 2. Balance the dataset without losing any existing training data --- Key AWS service reasoning Amazon SageMaker Clarify is the correct tool to analyze datasets for bias and compute metrics such as Class Imbalance (CI). It is specifically designed for data bias detection, not transformation. Amazon SageMaker Data Wrangler is used for data preprocessing and feature engineering, including resampling techniques like SMOTE. The constraint “without losing any existing training data” strongly indicates that undersampling should be avoided, because it removes majority-class samples. Therefore, the correct balancing method is oversampling (SMOTE), which synthetically increases the minority class instead of discarding data. --- Option analysis A) Clarify + CI = 0 → random undersampling Incorrect logic: If CI = 0, data is already balanced, so no action is needed. Uses undersampling, which violates the requirement of not losing data. Undersampling is only suitable when dataset is very large and loss of majority data is acceptable (not the case here). --- B) Clarify + CI > 0 → SMOTE in Data Wrangler Correct approach: Clarify detects imbalance using CI metric. If CI > 0, dataset is i...

Author: Noah · Last updated Aug 16, 2026

A healthcare company uses an Amazon SageMaker AI endpoint to host a model that predicts patient readmission risk to hospitals. The company wants to predict patient readmissions with high accuracy and is willing to tolerate false positives. The current model performance has degraded over the previous year. The company trains and deploys a new model as a shadow variant for testing on live traffic from hospitals. The company monitors the perf...

Key constraints in this question: The model is used for patient readmission risk prediction (high-impact healthcare workload). The business tolerates false positives, so recall is more important than precision. The shadow test shows: ✅ Higher recall (good, aligns with goal) ❌ Lower precision (more false positives, but acceptable per requirement) The model has only been tested in shadow mode for a month, meaning it has not yet been safely promoted to production traffic control. Option analysis A) Promote the shadow variant to full production ❌ Even though recall improved (the key metric), this is too abrupt. Shadow testing does not guarantee safe production behavior under controlled rollout conditions. In healthcare workloads, AWS best practice avoids direct full-cutover when risks are not fully bounded. B) Extend the shadow testing period ❌ Unnecessary. The model has already been observed for a month, and the decision signal is clear. Extending does not address deployment strategy, only delay...

Author: Vikram · Last updated Aug 16, 2026

An ML engineer is using an Amazon SageMaker Studio notebook to train a neural network by creating an estimator. The estimator runs a Python training script that uses Distributed Data Parallel (DDP) on a single instance that has more than one GPU. The ML engineer discovers that the training script is underutilizing GPU resources. The ML...

Key idea in the question The ML engineer is already running a distributed training script (DDP on multi-GPU in a single instance) and the problem is underutilized GPU resources. The requirement is very specific: > “Identify the point in the training script where resource utilization can be optimized.” So we are not just asked to observe utilization—we need fine-grained profiling inside the training code to locate bottlenecks (data loading, forward/backward pass, synchronization, CPU-GPU transfer, etc.). That immediately points to a profiler, not monitoring or logging services. --- ✅ Correct Option: B) Use SageMaker Profiler annotations Why B is correct Amazon SageMaker Profiler is specifically designed to: Trace GPU/CPU utilization at the step and operator level Identify bottlenecks inside training scripts Provide timeline views of compute vs wait time Highlight issues like: Input pipeline starvation (DataLoader bottlenecks) GPU idle time Communication overhead in DDP (all-reduce delays) Inefficient batch sizing or kernel launches Why annotations matter By adding profiler annotations in the training script, the engineer can: Mark key regions (data loading, forward pass, backward pass) Correlate GPU utilization drops with specific code sections Generate a detailed execution timeline report 👉 This directly satisfies: > “identify the point in the training...

Author: Leah Davis · Last updated Aug 16, 2026

A company's ML engineer is creating a classification model. The ML engineer explores the dataset and notices a column that is named day_of_week. The column's data consists of the following values: Monday, Tuesday, Wednesday, Thursday, Friday, Saturda...

We are dealing with a categorical nominal feature: `day_of_week = {Monday, Tuesday, Wednesday, Thursday, Friday, Saturday, Sunday}` There is no inherent order or ranking between these values, so the encoding method must avoid introducing false ordinal relationships. --- ✅ Correct Approach: C) One-hot encoding Why One-hot encoding is correct One-hot encoding converts each category into a separate binary column (0/1): | day_Mon | day_Tue | ... | day_Sun | | ------- | ------- | --- | ------- | | 1 | 0 | ... | 0 | Key reasons: Preserves nominal nature (no ordering implied) Produces binary representation per category Works well for most ML algorithms used in AWS services (e.g., SageMaker linear models, tree-based models, etc.) Prevents model from interpreting artificial relationships (e.g., Sunday > Monday) When to use: Low to medium cardinality categorical features Nominal variables (colors, days, cities, product types) When algorithms cannot handle categorical variables directly --- ❌ Why other options are incorrect A) Binary encoding Converts ...

Author: Daniel · Last updated Aug 16, 2026

SNAPSHOT - An ML engineer needs to use Amazon SageMaker to develop an ML solution for a company. The solution will use streaming video from cameras to count the number of people who walk past the company's store every day. Select and order the steps from the following list to implement the first version of the algorithm. Each step should be selected one time. (Select and order three.) * Choose a built-in algorithm...

Author: Lucas · Last updated Aug 16, 2026

SNAPSHOT - A company needs to combine data from multiple sources. The company must use Amazon Redshift Serverless to query an AWS Glue Data Catalog database and underlying data that is stored in an Amazon S3 bucket. Select and order the correct steps from the following list to meet these requirements. Select each step one time or not at all. (Select and order three.) * Attach the IAM role to the Redshift cluster. * Attach the IAM role to the Redshift namespace. * Create an external database in Amazon Redshift to point to the Data Catalog schema. * Create an external schema in Amazon Redshift to poi...

Author: IceDragon2023 · Last updated Aug 16, 2026

A company is developing an internal cost-estimation tool that uses an ML model in Amazon SageMaker AI. Users upload high-resolution images to the tool. The model must process each image and predict the cost of the object in the image. The mo...

Correct answer: B The best solution is: B) Store the images in an Amazon S3 bucket. Deploy the model on Amazon SageMaker AI. Use an asynchronous inference strategy for model inference. Use an Amazon Simple Notification Service (Amazon SNS) topic to notify users. --- Why Option B is correct This scenario has three key requirements: 1. High-resolution image uploads 2. Per-image ML inference processing 3. User notification when processing completes 1. Suitable storage: Amazon S3 Using Amazon S3 is ideal because: It is designed for large object storage (like high-resolution images) Integrates natively with SageMaker Supports event-driven ML pipelines --- 2. Correct inference mode: Asynchronous inference Amazon SageMaker AI asynchronous inference is specifically designed for: Large payloads (like high-resolution images) Long-running inference tasks Decoupled request/response processing Key behavior: User uploads image → request is accepted immediately SageMaker processes it in the background Result is written to S3 output location Completion event can trigger notifications This directly matches the requirement: “process each image and predict cost” without blocking the user --- 3. Notification mechanism: SNS Amazon SNS is the best fit because: It is designed for event-based notifications Supports email, SMS, Lambda, HTTP endpoints, etc. Works naturally with SageMaker async completion events So when inference completes → SNS notifies the user. --- Why the other options are incorrect ❌ Option A (S3 + Batch Transform + SQS) Problems: Bat...

Author: Zara · Last updated Aug 16, 2026

An ML engineer is configuring auto scaling for an inference component of a model that runs behind an Amazon SageMaker AI endpoint. The ML engineer configures SageMaker AI auto scaling with a target tracking scaling policy set to 100 invocations per model per minute. The SageMaker AI endpoint scales appropriately during normal business hours. However, the ML engineer notices that at the start of each business day, there are zero instances available to handle requests, wh...

The key problem here is cold start / scale-to-zero delay at the start of the business day. Even though the endpoint scales correctly during business hours based on invocations (100 invocations per model per minute), it still ends up at zero instances before traffic begins, causing the first requests to be delayed. So the requirement is not about better responsiveness during load, but about ensuring at least one instance is already warm and available before traffic arrives. --- Key reasoning points Current policy: Target tracking on invocations per model per minute Issue: Scale-in reduces capacity to 0 instances during idle/off-hours Problem type: Cold start / zero-capacity bootstrap problem Required fix: Ensure minimum baseline capacity > 0 at business start time --- Option analysis ❌ A) Reduce cooldown + lifecycle hook Cooldown controls how quickly scaling actions can re-trigger. Lifecycle hooks are used for custom instance setup steps (bootstrapping, configuration, warmup scripts), not for guaranteeing capacity availability. Even with reduced cooldown, it does NOT guarantee instances exist at start of day. 👉 When useful: Complex initialization of instances (loading models, dependencies) Delayed readiness after instance launch 🚫 Not solving zero-capacity issue → incorrect. --- ❌ B) Change target metric to CPU utilization CPU-based scaling is common in EC2-based workloads. But for SageMaker inference, invocations-based metrics are the correct primary driver, especially for model endpoints. CPU utilization is: ...

Author: Leah · Last updated Aug 16, 2026

A travel company wants to create an ML model to recommend the next airport destination for its users. The company has collected millions of data records about user location, recent search history on the company's website, and 2,000 available airports. The data has several categorical features with a target column that is expected to have a high-dimensional sparse matrix. The company needs to use Amazon SageMaker AI built-...

The correct answer is: Selected option: C) Use the Factorization Machines algorithm to recommend the next airport destination. --- Why Factorization Machines (FM) is the right choice Amazon SageMaker built-in Factorization Machines (FM) algorithm is specifically designed for: High-dimensional sparse data Categorical features converted using one-hot encoding Recommendation and ranking problems Implicit feedback datasets (clicks, searches, user behavior history) In this scenario: User location + search history + airport IDs → extremely sparse feature space 2,000 airports → large categorical feature expansion after one-hot encoding Goal is recommendation → not forecasting or clustering FM works well because it: Learns pairwise feature interactions efficiently Handles sparse matrices without dense feature explosion Performs strongly in recommendation systems (collaborative filtering style problems) --- Why the other options are incorrect A) CatBoost CatBoost is excellent for categorical features in general ML tasks However, in SageMaker built-in algorithm context, FM is the more direct and purpose-built recommendation algorithm CatBoost is typically used for: Classification Regression Ranking (but not optimized for large sparse...

Author: Joseph · Last updated Aug 16, 2026

SNAPSHOT - A company uses Amazon SageMaker AI to support ML workflows such as model training and deployment. Select the correct registry from the following list to meet the requirements for each use case with the...

Author: Aarav · Last updated Aug 16, 2026

SNAPSHOT - An ML engineer needs to automate the rebuild and redeployment of an ML model. Updates will occur when changes are made to the model's code base. The ML engineer must use AWS services to configure a continuous integration and continuous delivery (CI/CD) pipeline for the rebuild and redeployment. Select and order the steps from the following list to configure the CI/CD pipeline. Each step should be selected one time. (Select and order three.) * Invoke Amazon...

Author: Ava · Last updated Aug 16, 2026

A retail company is creating an AI-powered assistant for customers. The company has a large body of documentation that the assistant needs to use for general inquiries. The company wants any responses about prices to use o...

Correct Answer: A Why Option A is correct The key requirement is selective retrieval based on document age (≤ 1 month) specifically for price-related queries. This is a classic Retrieval-Augmented Generation (RAG) filtering problem, where you must control which documents are retrieved before the LLM generates an answer. Amazon Q Business supports this through document attribute filters (metadata filtering). Key reasoning factors: Amazon Q Business can ingest enterprise documents and apply metadata-based retrieval filters You can tag documents with attributes like `created_date` A filter such as “only retrieve documents from the last 30 days when query relates to pricing” ensures compliance This enforces correctness at the retrieval layer, which is the right place to control freshness constraints 👉 This is exactly what the scenario requires: query-dependent, time-based filtering of knowledge sources --- Why other options are incorrect ❌ Option B: Source attribution citation filter Citations control how sources are shown, not which documents are retrieved It affects transparency, not retrieval logic So even if citations are limited, the model may still use outdated documents internally Therefore, it does not enforce freshness constraints 👉 Use case: when you want traceability or compliance reporting, no...

Author: Ethan · Last updated Aug 16, 2026

SNAPSHOT - An ML engineer needs to use Amazon SageMaker hyperparameter tuning to reduce the training time for an ML model. Select and order the correct steps from the following list to meet this requirement. Each step should be selected one time or not at all. (Select and order three.) * Choose Bayesian optimization and increase the number of parameters. * Choose Hyperband tuning and decrease the numb...

Author: Akash · Last updated Aug 16, 2026

A company runs its ML workflows on an on-premises Kubernetes cluster. The ML workflows include ML services that perform training and inferences for ML models. Each ML service runs from its own standalone Docker image. The company needs to perform a lift and shift from the on-premises Kubernetes cluster to an Am...

The key requirement here is a lift-and-shift migration of existing Kubernetes-based ML workloads from an on-prem cluster to Amazon EKS, with least operational overhead. That means: Keep existing containerized workloads unchanged Avoid redesigning ML architecture Minimize changes to training/inference pipelines Focus on moving “as-is” workloads to AWS-managed Kubernetes --- ✅ Selected Option: B B) Upload the Docker images to Amazon Elastic Container Registry (Amazon ECR). Configure a deployment pipeline to deploy the images to the EKS cluster. Why this is correct This is the true lift-and-shift approach: Your ML services are already packaged as standalone Docker images You only need to move them to AWS infrastructure: Store images in Amazon Elastic Container Registry Deploy them on Amazon Elastic Kubernetes Service No application redesign required Minimal operational change (mostly CI/CD + cluster provisioning) Key factors: No code changes required Preserves existing Kubernetes manifests Direct container portability EKS handles orchestration; ECR handles image storage Lowest operational overhead among all options When this option is used: Migrating existing Docker/Kubernetes workloads to AWS “Lift-and-shift” container migration Keeping current ML training/inference services unchanged --- ❌ Why other options are wrong A) Redesign using Kubeflow Uses Kubeflow Requires major redesign of M...

Author: SilverBear · Last updated Aug 16, 2026

A company uses an Amazon QuickSight dashboard to track the sale prices of sneakers over time. The dashboard aggregates sale prices scraped from many retail websites. The company wants to determine which prices are unusually hi...

Key requirement breakdown Data is in Amazon QuickSight Goal is to identify unusually high price outliers Must visually display outliers Needs a built-in or analytics-friendly approach, not unnecessary external processing --- Correct approach: C Why C is correct Uses Amazon QuickSight anomaly detection insights, which is specifically designed for: Detecting statistical outliers in time series or numeric datasets Highlighting unusual spikes (high prices in this case) Keeps processing inside QuickSight, which is optimal since the dashboard already exists there Uses a bar chart visualization, which can clearly show anomalies when combined with insights Even though the “square transformation” is not meaningful analytically, the key correct component is: > QuickSight anomaly detection insights → correct AWS-native solution for outliers When to use this approach Use QuickSight anomaly detection when: You want to detect unexpected spikes or drops You are working with time-series or numeric business metrics You want visual overlays of anomalies in dashboards You want a no-ETL, built-in ML-based approach --- Why other options are incorrect ❌ A (Lambda + square root transformation) Square root transformation does not detect outliers It ...

Author: NightmareDragon2025 · Last updated Aug 16, 2026

A company wants to migrate ML models from an on-premises environment to Amazon SageMaker AI. The models are based on the PyTorch algorithm. The company needs to reuse its existing custom scripts as much as possible...

The correct answer is D) SageMaker AI script mode. Why Script Mode is correct Amazon SageMaker AI Script Mode allows you to bring your existing machine learning code (such as PyTorch training and inference scripts) and run it on managed SageMaker infrastructure with minimal changes. Key reasons it fits the requirement: Maximum code reuse: You can directly reuse existing PyTorch training scripts from on-premises. Framework support: Native support for frameworks like PyTorch, TensorFlow, and Scikit-learn. Minimal refactoring: Only small adjustments may be needed (e.g., input/output paths, environment variables). Managed infrastructure: SageMaker handles provisioning, scaling, and training orchestration. This exactly matches the requirement of migrating models while preserving custom scripts. --- Why other options are incorrect A) SageMaker AI built-in algorithms These are pre-implemented algorithms (e.g., XGBoost, Linear Learner). They require data to be adapted to SageMaker’s expected format a...

Author: NightmareDragon2025 · Last updated Aug 16, 2026

A company uses an NFS-based data store to store data for ML training. Linux-based systems access the data store. The company needs a hybrid system to make the shared data store accessible to on-premises servers and Amazon SageMaker AI notebooks that will consume th...

This is fundamentally a shared POSIX-compatible file system requirement with hybrid (on-prem + AWS) access and file locking for producers. Key requirement breakdown Existing system is NFS-based → implies need for POSIX-compliant shared file system Must support: On-premises Linux servers AWS ML environment (Amazon SageMaker AI notebooks) File locking required → important for concurrent writers/producers Hybrid access → must work across AWS + on-prem reliably --- Option analysis A) Amazon S3 + Mountpoint for S3 S3 is object storage, not a true file system Mountpoint for S3 provides a file-like interface, but: ❌ No true POSIX file locking semantics ❌ Not suitable for concurrent writers needing locking guarantees ❌ Eventual consistency model (not ideal for training data producers) Use case: best for data lakes, scalable read-heavy ML training datasets, not shared POSIX workloads ➡️ Rejected due to lack of real file locking and POSIX semantics --- B) Amazon Elastic File System (EFS) Fully managed NFS-based shared file system ✔ POSIX-compliant with file locking support ✔ Works with Linux systems natively ✔ Supports mounting from: AWS services (including SageMaker notebooks) On-prem servers via VPN / Direct Connect ✔ Designed for shared concurrent access wor...

Author: MoonlitPantherX · Last updated Aug 16, 2026