AWS Machine Learning Certification Exam Dumps and Braindumps

Which modeling approach enables cost efficient predictive maintenance that adapts to changing conditions over time?

• ❏ A. Deep neural network with many layers

• ❏ B. Random Forest with curated features and periodic retraining

• ❏ C. SageMaker XGBoost trained once with no retraining

• ❏ D. Amazon Timestream

• ❏ E. Logistic regression

Why choose Amazon Neptune for storing and querying highly connected data with fast relationship traversals?

• ❏ A. Optimized for complex SQL joins on relational tables

• ❏ B. Best for large-scale unstructured text ingestion with minimal ops

• ❏ C. Purpose-built graph database with nodes/edges and fast relationship traversals

• ❏ D. DynamoDB key-value access is better for graph traversals

For S3 encryption requiring fine-grained key access control, automatic rotation every 12 months, and full audit logs, which AWS KMS approach should be used?

• ❏ A. AWS Secrets Manager with IAM and CloudTrail

• ❏ B. KMS customer managed keys with key policies, IAM, 12-month rotation, and CloudTrail

• ❏ C. SSE-S3 with bucket policies and CloudTrail

• ❏ D. AWS managed keys with bucket policies and CloudTrail

Which setup ensures SageMaker training and processing use only private networking and enforce granular S3 access with no internet path?

• ❏ A. Route SageMaker traffic to S3 through a NAT Gateway

• ❏ B. Run SageMaker in private subnets with an S3 VPC endpoint plus endpoint and bucket policies

• ❏ C. Use public S3 endpoints secured with TLS

• ❏ D. Create interface endpoints for SageMaker APIs only and keep S3 public

In CNN training, validation accuracy drops after about 36 epochs while training accuracy continues to rise; which actions reduce overfitting and maintain validation performance? (Choose 2)

• ❏ A. Amazon SageMaker Clarify

• ❏ B. Use early stopping on validation loss

• ❏ C. Increase model depth and width

• ❏ D. Amazon SageMaker Automatic Model Tuning

• ❏ E. Add dropout layers for regularization

Which mapping of Spot, On-Demand, and Reserved Instances best balances cost, resiliency, and latency for real-time inference, 30-day batch retraining, and large hyperparameter tuning? (Choose 3)

• ❏ A. Hyperparameter tuning on Spot

• ❏ B. Real-time endpoints on Spot

• ❏ C. 30-day retraining on Reserved Instances

• ❏ D. Real-time inference on On-Demand

• ❏ E. Hyperparameter tuning on Reserved Instances

In Amazon QuickSight, which feature lets viewers choose values at runtime (for example, territory and product line) to dynamically filter visuals?

• ❏ A. Row-level security

• ❏ B. Parameters with controls

• ❏ C. Drill-through actions

• ❏ D. Drill-down hierarchy

What primary capability does Amazon Macie provide for S3 data to reduce PII exposure risk before model training?

• ❏ A. S3 encryption with KMS keys

• ❏ B. Automatic discovery and classification of sensitive data

• ❏ C. Built-in masking or anonymization

• ❏ D. Real-time VPC flow inspection

• ❏ E. AWS Lake Formation fine-grained access control

Which AWS deployment fits low-latency bursty recommendations, event-driven fraud detection, and nightly batch forecasting at 02:00 UTC?

• ❏ A. SageMaker Asynchronous Inference for recommender, Kinesis + Lambda for fraud, AWS Batch for forecasting

• ❏ B. Amazon EKS for recommender, SageMaker real-time endpoints for fraud, AWS Batch for forecasting

• ❏ C. SageMaker real-time endpoint for recommender, AWS Lambda for fraud events, Amazon ECS scheduled tasks for forecasting

• ❏ D. SageMaker Serverless Inference for recommender, AWS Batch for fraud, SageMaker Processing for forecasting

Which fully managed AWS service best extracts unique key phrases and custom entities from about 150,000 text documents per month?

• ❏ A. Amazon OpenSearch Service

• ❏ B. Amazon Comprehend with custom entity recognition

• ❏ C. Amazon Kendra

• ❏ D. Amazon Bedrock

All questions come from my Udemy AWS ML course, and certificationexams.pro

Which AWS approach enables scalable near-real-time SQL analytics on a streaming ingest with minimal data loss?

• ❏ A. Amazon Kinesis Data Firehose to Amazon Redshift

• ❏ B. Amazon MSK to Amazon OpenSearch Service

• ❏ C. Use Amazon Kinesis Data Streams with Kinesis Data Analytics (SQL)

• ❏ D. Amazon Kinesis Data Firehose to Amazon S3 and query with Amazon Athena

How should S3 data be partitioned and lifecycle-managed to optimize Athena queries over the latest 15 days while archiving after 90 days?

• ❏ A. Bundle each day into one large gzipped file and query 15-day trends

• ❏ B. Convert to Parquet with Snappy but keep data unpartitioned and filter for 15 days

• ❏ C. Partition by year=YYYY/month=MM/day=DD and transition to S3 Glacier after 90 days

• ❏ D. Store data in S3 Glacier Flexible Retrieval from day 1 and query with Athena after restore

Which configurations best meet strong encryption, least-privilege access, and auditability for a batch Amazon Comprehend job processing text in Amazon S3? (Choose 3)

• ❏ A. Attach AmazonS3FullAccess to the execution role

• ❏ B. Turn on CloudTrail and publish Comprehend metrics and logs to CloudWatch

• ❏ C. Skip encrypting the output

• ❏ D. Enforce TLS in transit and KMS encryption for input, output, and logs

• ❏ E. IAM role scoped to only required S3 prefixes and log resources

• ❏ F. Use SSE-S3 instead of KMS to simplify setup

For an inference API with under 120 requests most days but spikes to 6,000 within minutes, which SageMaker option auto-scales for bursts and avoids idle instance cost?

• ❏ A. Amazon SageMaker Asynchronous Inference

• ❏ B. Amazon SageMaker Serverless Endpoints

• ❏ C. Amazon SageMaker Real-Time Inference

• ❏ D. Amazon SageMaker Multi-Model Endpoints

For highly imbalanced (~0.5% positive) real-time fraud classification on tabular features using SageMaker built-in algorithms, which algorithm best handles imbalance and achieves strong accuracy?

• ❏ A. Random Cut Forest

• ❏ B. XGBoost with scale_pos_weight and F1 tuning

• ❏ C. Linear Learner with class weights

• ❏ D. IP Insights

Which AWS service provides managed, low-latency real-time ML inference with autoscaling and high availability?

• ❏ A. Amazon ECS

• ❏ B. Amazon SageMaker Endpoints

• ❏ C. AWS Lambda

• ❏ D. AWS App Runner

• ❏ E. Amazon EC2 Auto Scaling

A 25 TB image dataset is on Amazon FSx for NetApp ONTAP in the same VPC as SageMaker; what is the most efficient way for the training job to access it?

• ❏ A. Migrate to Amazon FSx for Lustre and use a file system channel

• ❏ B. Copy to Amazon S3 and read from S3

• ❏ C. Use SageMaker Pipe mode from S3 after syncing the data

• ❏ D. Mount the FSx for ONTAP NFS export to the SageMaker training job

How can you automatically discover and remove PII in about 25 TB of CSV and JSON data in Amazon S3 with minimal operations?

• ❏ A. AWS Glue DataBrew jobs for PII scan and cleanup

• ❏ B. Amazon S3 Object Lambda to redact PII on access

• ❏ C. Amazon Macie with EventBridge and AWS Lambda for redaction

• ❏ D. Amazon Comprehend PII detection across S3 objects

For large-scale distributed deep learning and real-time inference under 8 ms, which AWS training and inference setup is most performant and cost-efficient?

• ❏ A. SageMaker trn1 for training; ml.g5 for inference

• ❏ B. SageMaker p4d for training; ml.inf1 for inference

• ❏ C. EC2 g4dn for both training and inference

• ❏ D. SageMaker p3 for training; ml.m6i for inference

In SageMaker multi-node GPU training, which placement and data-locality actions reduce inter-node latency and improve throughput? (Choose 2)

• ❏ A. Use an EC2 spread placement group

• ❏ B. Co-locate training data and compute in the same Region and AZ

• ❏ C. Place all training instances in one subnet and AZ

• ❏ D. Keep training data in a different Region

• ❏ E. Spread instances across multiple subnets

Which approach enables scalable feature engineering and SageMaker training on 50 million DynamoDB items while avoiding high-latency point reads from the source table?

• ❏ A. Query DynamoDB via Athena federated connector and read with Athena

• ❏ B. Run an AWS Glue ETL job to write DynamoDB data to Amazon S3 and train from S3

• ❏ C. Use Boto3 from a SageMaker notebook to read items on demand during preprocessing

• ❏ D. Use SageMaker Pipe mode to stream directly from DynamoDB into the training job

Which AWS service monitors Amazon Kendra indexing metrics and lets you set alarms that trigger notifications?

• ❏ A. Amazon SNS

• ❏ B. AWS CloudTrail

• ❏ C. Amazon EventBridge

• ❏ D. Amazon CloudWatch Alarms

• ❏ E. AWS Lambda

How can you alert when a SageMaker real-time endpoint’s Invocations exceed a threshold during 8 p.m.–11 p.m.?

• ❏ A. Use AWS CloudTrail Insights to alert on API volume

• ❏ B. Route SageMaker events to Amazon EventBridge and trigger on a count threshold

• ❏ C. Create a CloudWatch Logs metric filter from endpoint logs and alarm on matches

• ❏ D. Set a CloudWatch alarm on the SageMaker endpoint Invocations metric with SNS notification

For a regression model predicting a continuous target, which metric reports the average absolute error in the target’s units?

• ❏ A. Amazon SageMaker Clarify

• ❏ B. Area Under the ROC Curve (AUC-ROC)

• ❏ C. Mean Absolute Error (MAE)

• ❏ D. Root Mean Squared Error (RMSE)

Which AWS services best provide low-latency, stateful 30-second window aggregations over Kinesis Data Streams with minimal operations? (Choose 2)

• ❏ A. AWS Lambda

• ❏ B. Amazon Managed Service for Apache Flink

• ❏ C. Amazon EMR with Apache Spark Streaming

• ❏ D. Amazon Kinesis Data Analytics

• ❏ E. Amazon MSK

• ❏ F. AWS Glue streaming jobs

The best choice is Random Forest with curated features and periodic retraining. Random Forests handle non-linear relationships and noisy sensor signals well, train efficiently compared to deep nets, and are straightforward to refresh on new data. Periodic retraining enables the model to adapt to concept drift as operating conditions evolve, maintaining accuracy while controlling costs.

The option Deep neural network with many layers is typically compute-intensive, requires extensive tuning, and is harder to operate under strict budget constraints.

Logistic regression is inexpensive but often underfits complex failure dynamics, leading to poor recall on rare events.

Amazon Timestream is a time series database for storage and queries, not a modeling technique, so it does not satisfy the predictive requirement.

SageMaker XGBoost trained once with no retraining locks the model to an initial data distribution; without scheduled updates it cannot adapt to drift, which is critical in predictive maintenance.

Purpose-built graph database with nodes/edges and fast relationship traversals is correct because Amazon Neptune is designed specifically for graph data models, representing entities and relationships as nodes and edges. It supports Gremlin and SPARQL for efficient multi-hop traversals and low-latency queries on highly connected datasets, which is ideal for relationship-centric workloads.

The option Optimized for complex SQL joins on relational tables is incorrect because Neptune is not a relational database and does not use SQL; relational joins suggest Amazon RDS or Amazon Aurora.

The option Best for large-scale unstructured text ingestion with minimal ops is incorrect because unstructured text search and analytics align more with Amazon OpenSearch Service or data lakes on Amazon S3.

The option DynamoDB key-value access is better for graph traversals is incorrect because while DynamoDB is excellent for key-based lookups, it is not optimized for complex, multi-hop relationship traversals common in graph workloads.

KMS customer managed keys with key policies, IAM, 12-month rotation, and CloudTrail is correct because customer managed KMS keys allow you to define precise key policies and use IAM and grants to tightly scope access, enable automatic annual rotation for symmetric keys, and record all KMS API usage in CloudTrail for compliance-grade auditing. For S3, use SSE-KMS pointing to this CMK to meet encryption and governance requirements.

AWS managed keys with bucket policies and CloudTrail is insufficient because AWS controls the key policy, preventing fine-grained, customer-defined permissions and governance.

SSE-S3 with bucket policies and CloudTrail relies on S3 managed keys, which do not support customer key policies or customer-controlled rotation.

AWS Secrets Manager with IAM and CloudTrail is designed for storing secrets, not for managing KMS keys, policies, or rotation used by S3 encryption.

Run SageMaker in private subnets with an S3 VPC endpoint plus endpoint and bucket policies is correct because placing SageMaker training and processing jobs in private subnets forces all egress through VPC-controlled paths, and using an Amazon S3 VPC endpoint (Gateway) keeps S3 access on the AWS network without internet traversal. Combining an endpoint policy with a bucket policy that restricts access to the specific VPC endpoint (for example via aws:SourceVpce) provides the granular, governance-grade controls required for confidential data.

Route SageMaker traffic to S3 through a NAT Gateway is wrong because a NAT Gateway uses the public internet, breaking the requirement for fully private networking.

Use public S3 endpoints secured with TLS is wrong because encryption in transit does not eliminate the use of internet paths.

Create interface endpoints for SageMaker APIs only and keep S3 public is wrong because privatizing SageMaker service APIs does not privatize S3 data-plane traffic; S3 would still be accessed via public endpoints.

Use early stopping on validation loss and Add dropout layers for regularization directly address the pattern of rising training accuracy with declining validation accuracy. Early stopping halts training at the point of best validation performance, preventing the model from continuing to memorize the training set. Dropout injects stochastic regularization by randomly zeroing activations during training, which curbs co-adaptation and improves generalization.

Amazon SageMaker Automatic Model Tuning can find good hyperparameters, but without explicit regularization or a stopping criterion it will not, by itself, stop overfitting within an individual training run.

Increase model depth and width raises capacity and typically intensifies overfitting when validation metrics are already degrading.

Amazon SageMaker Clarify addresses bias and explainability and does not mitigate overfitting during training.

The best mapping is Real-time inference on On-Demand, 30-day retraining on Reserved Instances, and Hyperparameter tuning on Spot. Real-time endpoints need immediate, reliable capacity with consistent latency, which On-Demand provides without interruption risk. A predictable monthly retraining baseline can commit to RIs to lock in discounts over a 1- or 3-year term. Hyperparameter tuning is batchy and fault-tolerant with checkpointing, making Spot the most cost-effective choice despite possible interruptions.

The option Real-time endpoints on Spot is unsuitable because interruptions can occur at short notice, harming availability and latency SLOs.

The option Hyperparameter tuning on Reserved Instances is inefficient because HPO usage is bursty and variable, so long-term commitments reduce flexibility and may waste reservations.

The correct choice is Parameters with controls. In QuickSight, parameters can be surfaced through dashboard controls (such as dropdowns or text boxes) so viewers set values at runtime. These parameter values can bind to filters across visuals, enabling dynamic, interactive filtering without redesigning visuals.

The option Row-level security is incorrect because it restricts data access by user but does not give viewers filter inputs.

The option Drill-through actions is incorrect since it enables navigation to another sheet or target with context applied rather than general, user-entered filtering across visuals.

The option Drill-down hierarchy is incorrect because it only changes the level of detail within a visual and does not provide a runtime control for selecting values.

Automatic discovery and classification of sensitive data is correct because Amazon Macie’s core function is to scan Amazon S3 and automatically identify, classify, and alert on sensitive data such as PII. This reduces exposure risk by enabling you to act on findings before downstream processing or training.

S3 encryption with KMS keys is incorrect because encryption at rest is provided by S3 and AWS KMS, not Macie.

Built-in masking or anonymization is incorrect since Macie does not redact or transform data; it only reports findings.

Real-time VPC flow inspection is unrelated to Macie and pertains to network telemetry and services like VPC Flow Logs or GuardDuty.

AWS Lake Formation fine-grained access control is incorrect because it governs permissions and table-level controls, not content discovery.

The best mapping is SageMaker real-time endpoint for recommender, AWS Lambda for fraud events, Amazon ECS scheduled tasks for forecasting. SageMaker real-time endpoints provide consistently low latency with autoscaling for bursty traffic, Lambda naturally handles event-driven invocation from payment events with spiky patterns, and ECS scheduled tasks (via EventBridge) are a straightforward fit for nightly batch forecasting.

The option SageMaker Asynchronous Inference for recommender, Kinesis + Lambda for fraud, AWS Batch for forecasting is suboptimal because Asynchronous Inference introduces queueing and higher response times, which conflicts with ultra-low-latency recommendations. The choice Amazon EKS for recommender, SageMaker real-time endpoints for fraud, AWS Batch for forecasting misaligns the event-driven fraud pattern with endpoints and does not maximize the managed ML serving benefits for the recommender.

The option SageMaker Serverless Inference for recommender, AWS Batch for fraud, SageMaker Processing for forecasting suffers from serverless cold starts for real-time spikes, Batch not being real-time for fraud, and Processing being aimed at data preparation rather than scheduled inference.

Amazon Comprehend with custom entity recognition is the right choice because it natively extracts key phrases at scale and allows training a custom entity recognizer for domain-specific entities, all as a fully managed service with low operational overhead. It supports batch processing through asynchronous APIs and scales to large document volumes with minimal maintenance.

The option Amazon OpenSearch Service is incorrect because it is a search and analytics engine that would require custom analyzers and ingestion pipelines and does not provide native semantic key phrase or custom entity extraction.

The option Amazon Kendra is incorrect because it is built for enterprise search and content retrieval, not for generating de-duplicated key phrase or entity outputs from text corpora.

The option Amazon Bedrock is incorrect because it focuses on foundation models and generative AI; while possible with custom prompts and orchestration, it is not a turnkey, managed solution for key phrase and custom entity extraction like Comprehend.

The correct choice is Use Amazon Kinesis Data Streams with Kinesis Data Analytics (SQL). Kinesis Data Streams provides durable, scalable ingestion with shard-based throughput and at-least-once delivery, minimizing data loss if producers or consumers restart. Kinesis Data Analytics (SQL) runs continuous SQL over the stream with low latency, enabling true near-real-time analytics.

The option Amazon Kinesis Data Firehose to Amazon Redshift is micro-batch oriented with buffering intervals, so it cannot deliver continuous streaming SQL over in-flight data.

The option Amazon MSK to Amazon OpenSearch Service targets search and log analytics, not streaming SQL; OpenSearch is optimized for indexing and search queries rather than continuous SQL processing.

The option Amazon Kinesis Data Firehose to Amazon S3 and query with Amazon Athena is also batch-based; Athena queries data at rest in S3, which introduces buffering and object commit delays unsuited to near-real-time streaming SQL.

Partition by year=YYYY/month=MM/day=DD and transition to S3 Glacier after 90 days is correct because date-based partitions allow Athena to prune to only the latest 15 days, minimizing scanned bytes and query cost, while an S3 Lifecycle transition to Glacier satisfies the archival requirement at lower storage cost.

The option Convert to Parquet with Snappy but keep data unpartitioned and filter for 15 days is wrong because without partitions, Athena still scans large portions of data even if the query filters by date. Parquet helps, but the lack of partition pruning keeps costs high.

The option Bundle each day into one large gzipped file and query 15-day trends is wrong since compression alone does not provide partition pruning, and very large files can still trigger broad scans. It also omits the required lifecycle archival step.

The option Store data in S3 Glacier Flexible Retrieval from day 1 and query with Athena after restore is incorrect because Athena cannot directly read Glacier storage classes; objects must be restored to S3 first, introducing delays and operational overhead.

Enforce TLS in transit and KMS encryption for input, output, and logs, IAM role scoped to only required S3 prefixes and log resources, and Turn on CloudTrail and publish Comprehend metrics and logs to CloudWatch together provide defense in depth for data confidentiality, access control, and auditability. Comprehend supports using KMS keys for input and output data, TLS protects data in transit, least-privilege IAM restricts the blast radius, and CloudTrail plus CloudWatch deliver traceability and monitoring needed for compliance.

The option Attach AmazonS3FullAccess to the execution role is overly permissive and violates least-privilege, increasing risk of unintended access. The choice Skip encrypting the output is unsafe because output can contain sensitive derived content and must be encrypted. The distractor Use SSE-S3 instead of KMS to simplify setup may encrypt at rest but lacks the key management and control typically required by strong security policies, making it insufficient for strict compliance.

The correct choice is Amazon SageMaker Serverless Endpoints. It automatically scales concurrency based on incoming requests and charges only for compute used during invocation, eliminating idle instance costs. This makes it ideal for unpredictable, bursty traffic where you want low operational overhead and no manual capacity planning.

Amazon SageMaker Asynchronous Inference is optimized for long-running or large payload jobs where responses can be retrieved later, not for low-latency interactive APIs during sudden bursts.

Amazon SageMaker Real-Time Inference relies on provisioned instances that accrue cost even when idle, which conflicts with the goal of minimizing idle cost for sporadic traffic.

Amazon SageMaker Multi-Model Endpoints can consolidate multiple models on shared instances to reduce costs, but they still require provisioned instances and therefore do not eliminate idle capacity charges or management.

The best choice is XGBoost with scale_pos_weight and F1 tuning. Gradient-boosted trees are state-of-the-art for tabular, heterogeneous features and support direct class-imbalance handling via scale_pos_weight, enabling tuning toward precision, recall, and F1 for rare-event fraud detection. With labels available, supervised boosting will typically outperform unsupervised anomaly detectors and linear baselines.

Random Cut Forest is incorrect because it is unsupervised and does not leverage labels to optimize supervised metrics like precision/recall for known fraud classes.

Linear Learner with class weights is not ideal because linear models miss nonlinear feature interactions and generally underperform boosted trees on complex fraud patterns even with weighting.

IP Insights is designed for detecting unusual entity-to-IP relationships and is unsupervised, so it is not a fit for supervised transaction classification with labels. On the exam, when you see imbalanced labeled tabular data, prefer boosted-tree methods in SageMaker. Look for cues like scale_pos_weight or tuning for precision/recall/F1, and avoid unsupervised algorithms when labels are available.

All questions come from my Udemy AWS ML course, and certificationexams.pro

Amazon SageMaker Endpoints is correct because it offers fully managed real-time inference with automatic scaling via Application Auto Scaling, multi-AZ high availability, integrated monitoring, and deployment features purpose-built for production ML workloads.

The option Amazon ECS is incorrect because, while it can host containers, you must implement the model server, autoscaling, health checks, and multi-AZ strategies yourself, increasing operational complexity.

The option AWS Lambda is incorrect due to cold starts, timeouts, and concurrency constraints that make it less suitable for consistent, low-latency, high-throughput inference.

The option AWS App Runner is incorrect because it runs and scales containers but lacks ML-native inference features like model endpoints, variants, and built-in metrics for model serving.

The option Amazon EC2 Auto Scaling is incorrect since it only scales instances and does not provide a managed inference endpoint or deployment workflow.

Mount the FSx for ONTAP NFS export to the SageMaker training job is correct because the data is already on a high-throughput network file system within the same VPC. Mounting and reading in place avoids any bulk transfers, minimizes latency, and maximizes throughput for large image datasets during training.

The option Migrate to Amazon FSx for Lustre and use a file system channel is unnecessary because it introduces a full data migration step even though Lustre is performant.

The option Copy to Amazon S3 and read from S3 adds time, cost, and S3 request latency, which is less efficient for many files compared to direct NFS reads.

The option Use SageMaker Pipe mode from S3 after syncing the data still requires moving data to S3 and may be throughput limited versus in-VPC NFS, making it suboptimal when in-place access is available.

Amazon Macie with EventBridge and AWS Lambda for redaction is correct because Macie is a managed service purpose built for discovering sensitive data in Amazon S3. It natively generates findings for PII and integrates with EventBridge so you can invoke Lambda to automatically redact or quarantine objects, achieving minimal ongoing operations while maintaining compliance and automation.

AWS Glue DataBrew jobs for PII scan and cleanup is not ideal because you must build and maintain recipes, schedules, and scaling, which adds operational overhead versus a managed PII discovery service.

Amazon S3 Object Lambda to redact PII on access is not sufficient because it still requires you to implement custom PII detection and transformation logic and does not provide automated discovery across at-rest S3 data.

Amazon Comprehend PII detection across S3 objects would require custom orchestration to crawl and process large S3 datasets and manage retries and scaling, leading to higher complexity and operations.

SageMaker p4d for training; ml.inf1 for inference is the best match because p4d provides A100 GPUs for highly scalable distributed training, while AWS Inferentia on ml.inf1 is purpose-built for high-throughput, low-latency, and cost-efficient real-time inference. This combination aligns with both the performance and cost goals for sub-8 ms serving. The SageMaker trn1 for training; ml.g5 for inference choice can train effectively, but GPU inference on g5 is generally costlier per throughput and not as latency-optimized as Inferentia for large fleets. The EC2 g4dn for both training and inference setup is inadequate for massive training scale and struggles to hit ultra-low latency at high throughput. The SageMaker p3 for training; ml.m6i for inference choice pairs older training hardware with CPU inference, which is unlikely to meet strict millisecond latency for deep models. On the exam, map distributed training and scale to A100-class instances such as p4d (or p5) and map single-digit ms, cost-efficient inference to Inferentia (Inf1/Inf2). Watch for distractors that use CPUs or general-purpose GPUs for inference when low latency and cost efficiency are emphasized.

The best choices are to keep communication paths as short and local as possible. Placing the compute fleet together and keeping the data close reduces hops, routing, and cross-AZ or cross-Region penalties. Therefore, Co-locate training data and compute in the same Region and AZ and Place all training instances in one subnet and AZ together minimize inter-node latency and improve throughput in tight, collective communication patterns common in distributed training.

The option Use an EC2 spread placement group focuses on failure isolation, distributing instances across distinct hardware, which works against node adjacency and can increase latency.

The option Keep training data in a different Region incurs high latency and egress costs, which is counterproductive for frequent parameter synchronization.

The option Spread instances across multiple subnets can add routing overhead and jitter, increasing inter-node latency rather than reducing it.

The best choice is Run an AWS Glue ETL job to write DynamoDB data to Amazon S3 and train from S3. Moving data to S3 provides a durable, scalable data lake that SageMaker can read in parallel with high throughput. This avoids high-latency, per-item reads from DynamoDB and decouples your training jobs from the transactional workload. Glue offers managed, scalable ETL for bulk export and transformation, which is ideal for feature engineering at this scale.

The option Query DynamoDB via Athena federated connector and read with Athena is suboptimal for sustained training I/O because federated queries introduce Lambda invocation overhead and are intended for interactive analytics, not continuous, large-scale data feeds.

The option Use Boto3 from a SageMaker notebook to read items on demand during preprocessing relies on point reads which cause high latency and potential throttling, making it unreliable and slow for tens of millions of items.

The option Use SageMaker Pipe mode to stream directly from DynamoDB into the training job is not supported; Pipe mode streams from S3 or supported file systems, and even if it did, it would still place unnecessary load on DynamoDB.

Amazon CloudWatch Alarms is correct because Amazon Kendra publishes operational metrics and logs to CloudWatch, where you can create alarms on error or failure metrics and trigger notifications (for example, via Amazon SNS) or automated actions.

The option Amazon SNS is incorrect because it only delivers notifications and requires another service like CloudWatch to detect the error condition.

Amazon EventBridge is incorrect because it focuses on event routing and rules, not on collecting Kendra metrics or creating metric-based alarms.

AWS Lambda is incorrect since it is compute for running code and does not provide native monitoring or alarms.

AWS CloudTrail is incorrect because it logs API activity, not service metrics or alarms for indexing jobs.

The correct approach is to use Set a CloudWatch alarm on the SageMaker endpoint Invocations metric with SNS notification. SageMaker real-time endpoints publish AWS/SageMaker metrics such as Invocations and Invocation4XXErrors to CloudWatch. You can create a threshold alarm on Invocations and configure an SNS action for alerts. To limit alerts to specific hours (8 p.m.–11 p.m.), enable or disable the alarm on a schedule using EventBridge, or use composite alarms to gate notifications during that window.

The option Use AWS CloudTrail Insights to alert on API volume is incorrect because CloudTrail primarily captures management-plane activity and Insights detects anomalies in those events; it does not monitor data-plane invoke counts for endpoints.

The option Route SageMaker events to Amazon EventBridge and trigger on a count threshold is incorrect since EventBridge lacks built-in metric aggregation and threshold alarms; you would need custom aggregation.

The option Create a CloudWatch Logs metric filter from endpoint logs and alarm on matches is unreliable because endpoints do not produce access logs by default and log-based counting is indirect and brittle compared to native metrics.

Mean Absolute Error (MAE) is correct because it directly averages the absolute differences between predictions and ground truth and reports the result in the target’s units, providing an intuitive measure of typical error magnitude for regression.

Area Under the ROC Curve (AUC-ROC) is a classification ranking metric and does not apply to continuous regression targets.

Amazon SageMaker Clarify is a service for bias detection and explainability, not an evaluation metric.

Root Mean Squared Error (RMSE) is a valid regression metric and shares the target’s units, but it computes the square root of the mean squared error, emphasizing larger errors rather than reporting an average absolute error as asked.

Amazon Managed Service for Apache Flink and Amazon Kinesis Data Analytics are purpose-built for stateful, low-latency stream processing with windowed aggregations over Kinesis Data Streams. Flink provides rich state management and event-time windows with exactly-once semantics, while Kinesis Data Analytics enables concise SQL window functions to compute rolling metrics with minimal operational overhead. Both meet the 30-second rolling window and low-latency requirement without managing clusters.

The option AWS Lambda is not ideal because maintaining cross-shard, stateful 30-second windows is complex, error-prone, and can hit scaling or state limits.

Amazon EMR with Apache Spark Streaming can do windows but increases operational burden and latency due to cluster management and tuning.

Amazon MSK is a managed Kafka ingestion service, not an analytics engine for Kinesis streams, and does not solve windowed aggregation needs here.

AWS Glue streaming jobs focus on micro-batch ETL and generally do not deliver the lowest-latency rolling aggregations compared to Flink or Kinesis Data Analytics.