AWS MLA-C01 Practice Questions with Explanations

Correct answer: C

SageMaker Model Registry with model groups is the purpose-built AWS service for managing model versions and metadata with minimal operational overhead. It provides native integration with SageMaker's ML lifecycle, automatic versioning, and approval workflows. ECR repositories are container registries not designed for model metadata management, and tagging alone (options B and D) lacks the governance and cataloging features of the Model Registry.

Why the other options are wrong:

• A. ECR repositories are for container images, not optimized for ML model versioning and governance.
• B. ECR with tags requires manual management and lacks SageMaker's model-specific features like approval status and lineage tracking.
• D. Generic tagging in Model Registry lacks the organizational structure that model groups provide for managing related model versions.

Correct answer: B

SageMaker managed warm pools maintain pre-initialized compute resources between training jobs, eliminating the startup overhead of provisioning and initializing instances. This directly addresses minimizing infrastructure startup times for consecutive jobs. Managed Spot Training reduces costs but not startup time. Training Compiler optimizes code execution speed, not infrastructure provisioning. SMDDP is for distributed training parallelism, not startup optimization.

Why the other options are wrong:

• A. Managed Spot Training reduces costs through interruption-prone instances but does not minimize startup times.
• C. Training Compiler optimizes training job execution speed, not the infrastructure provisioning startup phase.
• D. Distributed data parallelism is for scaling training across multiple instances, not reducing startup times.

Correct answer: D

SageMaker Pipelines enables automated workflows that can programmatically manage model approval status through the AWS SDK. When a model is registered in the Model Registry, the pipeline can conditionally change the approval status to 'Approved' based on validation checks, creating a manual approval gate for production deployment. SageMaker Experiments tracks experiments but doesn't manage approvals. ML Lineage Tracking records relationships but doesn't enforce approval gates. Model Monitor evaluates performance but is not designed for approval workflows.

Why the other options are wrong:

• A. SageMaker Experiments facilitates experiment organization and comparison, not approval process management.
• B. ML Lineage Tracking documents model relationships and data provenance but does not implement approval workflows.
• C. Model Monitor detects model performance degradation and data drift but is not designed for approval-based gating.

Correct answer: A

SageMaker Clarify is the dedicated service for bias detection and monitoring, and invoking it through AWS Lambda enables on-demand workflow execution for monitoring bias drift in deployed endpoints. This provides the required monitoring capability without continuous execution overhead. Pulling the monitor-analyzer image manually (option B) bypasses SageMaker's integrated monitoring. AWS Glue Data Quality monitors data quality but not model bias. Notebooks require manual execution and don't provide on-demand automation.

Why the other options are wrong:

• B. Manually pulling the monitor-analyzer image requires custom scripting and lacks the integration that Clarify provides.
• C. AWS Glue Data Quality monitors data quality issues, not model-specific bias drift in predictions.
• D. SageMaker notebooks require manual execution and comparison, not automated on- demand workflows.

Correct answer: D

AWS Lake Formation is purpose-built for aggregating and managing data from multiple sources (on-premises databases, S3, etc.) into a centralized data lake with unified governance and schema management. It provides connectors for on-premises databases like MySQL and integrates with S3 data. EMR is for processing but not aggregation governance. Kinesis Streams handle real-time streaming data, not batch aggregation from heterogeneous sources. DynamoDB is a NoSQL database, not an aggregation service.

Why the other options are wrong:

• A. EMR provides processing capability but not data aggregation and governance from multiple heterogeneous sources.
• B. Kinesis Data Streams is designed for real-time streaming data, not aggregating batch data from on-premises databases.
• C. DynamoDB is a NoSQL database service, not a data aggregation and governance platform for multiple source types.

Correct answer: C

SageMaker Data Wrangler includes built-in data quality and anomaly detection capabilities with integrated visualization features, enabling both anomaly detection and result visualization in a single tool. This meets both requirements with minimal additional components. Athena is a query service without anomaly detection. Redshift Spectrum querying data doesn't automatically detect anomalies; QuickSight visualization requires separate anomaly logic. AWS Batch requires custom anomaly detection code.

Why the other options are wrong:

• A. Amazon Athena is a query engine without built-in anomaly detection capabilities.
• B. Redshift Spectrum requires custom anomaly detection logic; it doesn't automatically detect anomalies without additional implementation.
• D. AWS Batch requires custom container code to detect anomalies and doesn't provide automatic anomaly detection.

Correct answer: C

SageMaker Data Wrangler provides visual, low-code data transformation with built-in recipes for encoding categorical variables into numerical representations. It requires minimal coding effort and provides immediate preview of transformations. AWS Glue requires writing custom transformation code (Spark/Python), increasing operational overhead. Data Wrangler's visual interface and pre-built transformations make it the least operationally intensive option.

Why the other options are wrong:

• A. AWS Glue requires custom ETL code or mapping logic, which increases operational complexity compared to Data Wrangler's visual interface.
• B. Converting numerical to categorical data is the inverse of what's needed for a classification model; it reduces rather than maximizes accuracy.
• D. Converting numerical to categorical data reverses the needed transformation and would degrade model performance.

Correct answer: D

SageMaker Data Wrangler includes a built-in balance data operation specifically designed to handle class imbalance through oversampling techniques. This requires no custom code and is the least operationally intensive solution. Athena is for querying, not data balancing. Studio Classic built-in algorithms may handle imbalance but require model training iteration rather than preprocessing. AWS Glue DataBrew is primarily for data quality and profiling, not specialized imbalance handling.

Why the other options are wrong:

• A. Athena identifies patterns through queries but doesn't provide automated balancing operations.
• B. SageMaker algorithms can handle imbalanced data but require training iterations rather than resolving the issue in preprocessing.
• C. AWS Glue DataBrew is designed for data quality and profiling, not providing built-in class balancing operations.

Correct answer: A

LightGBM is a gradient boosting algorithm that excels at handling class imbalance through built-in mechanisms like balanced bagging and feature-dependent penalties. It also captures complex feature interdependencies through its tree-based ensemble approach, making it ideal for capturing underlying patterns that simpler algorithms miss. The other options are unsuitable: Linear Learner cannot capture nonlinear interdependencies, K- means is unsupervised and not designed for classification, and NTM is for topic modeling on text data.

Why the other options are wrong:

• B. Linear Learner cannot capture nonlinear feature interdependencies and performs poorly with class imbalance.
• C. K-means is an unsupervised clustering algorithm, not appropriate for fraud detection classification.
• D. Neural Topic Model is designed for topic modeling on text corpora, not structured transactional fraud detection.

Correct answer: A

Concept drift occurs when the statistical properties of the target variable change over time in production data. After months of stable performance, a sudden significant decrease in F1 score indicates the underlying customer behavior patterns have shifted-customers likely differ in their cancellation patterns from the baseline period. This is the most common reason for model performance degradation in production. The other options describe static issues that would have affected performance from the beginning, not caused degradation after months of stability.

Why the other options are wrong:

• B. If the model was insufficiently complex for baseline data, performance would have been poor from the start, not degrade after stable months.
• C. Missing values in baseline data would have impacted initial model training and baseline performance, not cause degradation months later.
• D. Incorrect ground truth labels during baseline calculation would affect the threshold itself, not explain actual performance changes in production.

Correct answer: A

Creating a single IAM role with necessary permissions and attaching it to each notebook instance provides centralized permission management with minimal complexity. This approach allows the company to update permissions in one place (the role) and have those changes automatically apply to all notebook instances that use it. Option B fails because IAM groups cannot be directly associated with EC2 instances. Option C violates security best practices by using a shared IAM user instead of roles. Option D is overly complex and groups cannot be directly associated with instances.

Why the other options are wrong:

• B. IAM groups cannot be directly associated with or attached to notebook instances; only roles can be used.
• C. Using a shared IAM user with AdministratorAccess violates security best practices and creates accountability issues.
• D. This approach is unnecessarily complex; groups cannot be directly associated with notebook instances, only roles can.

Correct answer: D

Predicting apartment prices is a regression problem requiring a regression metric. Mean Absolute Error (MAE) measures the average absolute difference between predicted and actual prices in the same units as the target variable, making it directly interpretable and appropriate for price prediction. Accuracy, AUC, and F1 score are all classification metrics used for binary or multiclass prediction problems, not continuous value prediction.

Why the other options are wrong:

• A. Accuracy is a classification metric measuring the percentage of correct predictions, not applicable to regression.
• B. Area Under the ROC Curve is a classification metric for evaluating binary classifiers, not for regression tasks.
• C. F1 score is a classification metric combining precision and recall, not designed for regression problems.

Correct answer: D

The oscillating pattern of training and validation loss indicates the learning rate is too high, causing the optimizer to overshoot the minimum and bounce around it. Decreasing the learning rate allows SGD to take smaller, more controlled steps toward the optimal solution, stabilizing the training process. Early stopping (A) addresses overfitting, not oscillation. Increasing test set size (B) is irrelevant to training dynamics. Increasing learning rate (C) would worsen oscillation.

Why the other options are wrong:

• A. Early stopping addresses overfitting, not the oscillating loss pattern caused by high learning rate instability.
• B. Test set size does not affect training dynamics or the oscillating loss behavior during training.
• C. Increasing the learning rate would exacerbate the oscillation problem, not solve it.

Correct answer: A

Amazon Athena with CTAS statements provides serverless SQL querying directly against S3 CSV objects with minimal operational overhead. Athena automatically handles schema inference for CSV files and can partition/filter by transaction date without requiring data movement or complex job configurations. Option B adds unnecessary replication and S3 Object Lambda complexity. Option C requires setting up and managing Glue jobs and infrastructure. Option D requires Firehose setup and Lambda function development, both unnecessarily complex.

Why the other options are wrong:

• B. This solution adds unnecessary complexity with S3 replication and Object Lambda when simple querying suffices.
• C. AWS Glue requires job creation, management, and cluster configuration, adding operational overhead versus serverless Athena.
• D. Firehose and Lambda function development adds code complexity and operational overhead compared to Athena's serverless approach.

Correct answer: D

AWS Glue FindMatches is a purpose-built ML transform specifically designed to detect and remove duplicate records in datasets with minimal code. It uses machine learning to identify fuzzy matches across key attributes without requiring custom model development. Option A requires manual human review at scale. Option B (QuickSight) is for visualization and insights, not deduplication. Option C (Data Wrangler) is a general preprocessing tool requiring manual deduplication logic.

Why the other options are wrong:

• A. Mechanical Turk requires manual human review of records, which is not scalable for large datasets.
• B. QuickSight is a visualization and BI tool, not designed for duplicate detection functionality.
• C. Data Wrangler is a general preprocessing tool that would require custom logic development for deduplication.

Correct answer: A

Spot Instances offer up to 90% discount compared to On-Demand prices, making them the most cost-effective option for workloads that can tolerate interruptions. Since the batch job is scheduled for weekends and can handle interruptions, Spot Instances are ideal. Reserved Instances require upfront commitment and provide better economics for continuous workloads. On-Demand is full price. Dedicated Instances are for compliance requirements and are more expensive.

Why the other options are wrong:

• B. Reserved Instances require longer-term commitment and are less cost-effective than Spot for interruptible workloads.
• C. On-Demand pricing is full price, significantly more expensive than Spot for interrupt- tolerant workloads.
• D. Dedicated Instances are for compliance/licensing requirements and offer no cost advantage over standard instances.

Correct answer: B

Amazon Comprehend provides the ImportModel API operation specifically designed to copy custom models between AWS accounts in the same region. This is the native, purpose-built solution that requires minimal development effort. You create a resource- based IAM policy granting Account B permission to import the model from Account A, then use the ImportModel API. This is more efficient than manual S3 transfers, doesn't require DataSync infrastructure, and avoids the complexity of VPN setup.

Why the other options are wrong:

• A. Manual S3 copying doesn't leverage AWS's native model transfer capabilities and requires more manual steps.
• C. DataSync is designed for data synchronization workflows, not for copying Comprehend models.
• D. Site-to-Site VPN is unnecessarily complex for a cross-account operation within the same region and creates operational overhead.

Correct answer: C

Amazon Kendra is purpose-built for semantic search and can directly ingest documents from S3, automatically handling embeddings and vector indexing internally. It's a fully managed service that eliminates the need to manually generate embeddings, manage vector databases, or write custom queries. Options A and B require manual embedding generation and storage, losing the managed service benefits. Option D is incorrect because Amazon Textract is for text extraction, not semantic search.

Why the other options are wrong:

• A. AWS Glue is a data catalog and ETL service, not designed for vector storage or semantic search queries.
• B. This approach requires manual embedding generation and Feature Store doesn't provide semantic search capabilities-SQL queries cannot perform semantic similarity searches.
• D. Amazon Textract extracts text from documents but provides no semantic search or vector database capabilities.

Correct answer: A

Amazon SageMaker Autopilot is specifically designed to automatically build, train, and tune ML models with minimal user input. It can directly query Athena datasets, automatically handles feature engineering and model selection, and reports performance metrics. This requires the least development effort compared to implementing custom scripts (B), using Macie which is designed for data classification not model building (C), or fine-tuning foundation models from Bedrock which adds unnecessary complexity (D).

Why the other options are wrong:

• B. Custom scripts require significant development effort for preprocessing, model implementation, and evaluation.
• C. Amazon Macie is designed for data protection and sensitive data discovery, not for building predictive models.
• D. Bedrock foundation models are not designed for this use case and fine-tuning them adds unnecessary complexity and cost.

Correct answer: D

One-hot encoding (D) is the standard technique for handling categorical features in neural networks. It converts each category into a binary vector, preventing the model from incorrectly assuming ordinal relationships between color categories. Label encoding (A) assigns arbitrary integers, causing the model to interpret ordinal relationships that don't exist. Padding (B) addresses sequence length, not categorical encoding. Dimensionality reduction (C) is premature before encoding and would lose information.

Why the other options are wrong:

• A. Label encoding creates false ordinal relationships between colors (e.g., red=1, blue=2 implies blue > red).
• B. Padding ensures sequence length uniformity, which is irrelevant for non-sequential categorical color data.
• C. Dimensionality reduction should occur after proper encoding if needed, not as the initial categorical transformation.

Correct answer: C

Amazon Macie is a managed service that automatically identifies and classifies sensitive data in S3, requiring minimal operational overhead. Combined with Lambda functions for removal, this is fully serverless and requires no infrastructure management. Option A requires redeploying to SageMaker. Option B involves complex ECS/Fargate orchestration and Batch jobs. Option D requires manual EC2 instance management and Comprehend is designed for NLP tasks, not general sensitive data discovery.

Why the other options are wrong:

• A. Requires redeploying the on-premises model to SageMaker, changing the deployment architecture unnecessarily.
• B. ECS/Fargate and Batch require more operational overhead than using a managed service like Macie.
• D. EC2 instances require operational overhead. Comprehend is designed for NLP analysis, not general sensitive data identification.

Correct answer: B

AWS Glue is the purpose-built service for data ingestion and ETL pipelines from S3, providing native connectors, schema discovery, and transformation capabilities. SageMaker Studio Classic is designed for ML model deployment pipeline orchestration. Together they address both requirements with AWS's native ML stack. DataFirehose (A) is for streaming data. Redshift ML (C) is for SQL-based ML. Athena (D) is a query engine, not a pipeline builder.

Why the other options are wrong:

• A. Data Firehose is designed for streaming data ingestion, not batch data ingestion from S3 buckets.
• C. Redshift ML is for in-database ML, not for building general-purpose data ingestion pipelines.
• D. Athena is a query engine, not a pipeline orchestration tool. Using a notebook is less maintainable than SageMaker Studio.

Correct answer: D

SageMaker Model Registry Collections is the native feature designed to organize model groups hierarchically without affecting existing model artifacts or groupings. Collections allow organizing model groups by team or category while preserving their integrity. Option A uses tags but doesn't organize into the hierarchical team structure requested. Option B moves models between groups, disrupting existing groupings. Option C doesn't address the organizational need-lineage tracking is about model provenance, not categorization.

Why the other options are wrong:

• A. Tags provide metadata but don't create organizational hierarchies for improved discoverability at scale.
• B. Moving models to new category groups disrupts existing model group integrity and organization.
• C. ML Lineage Tracking records model dependencies and history, not organizational categorization by team.

Correct answer: B

Network ACLs are stateless and operate at the subnet level, making them ideal for blocking traffic from specific IP addresses at the network perimeter. A network ACL inbound deny rule will block all traffic from the specified IP address to any resource in the subnet, including the SageMaker domain. This is more comprehensive than security group rules, which are stateful and instance-level, and network ACLs are the proper mechanism for subnet-wide traffic filtering.

Why the other options are wrong:

• A. Security groups are stateful and work at the instance level; they cannot explicitly deny traffic (only allow or implicitly deny), and SageMaker domains don't directly attach security groups in the traditional sense.
• C. Shadow variants and Inference Recommender are for canary deployments and performance optimization, not for blocking malicious traffic.
• D. Route tables control packet routing between networks, not inbound traffic filtering; they cannot deny specific source IPs.

Correct answer: B

Amazon Transcribe converts audio and video to text efficiently, Amazon Translate handles Spanish-to-English conversion, and Amazon Bedrock provides instant access to Claude and other LLMs without training or deployment overhead. This fully managed, serverless approach minimizes time-to-value compared to training custom models or using Rekognition (for images only) or Comprehend (not designed for this workflow).

Why the other options are wrong:

• A. Training and deploying custom SageMaker models requires significant time for model development, training, and endpoint deployment.
• C. Rekognition is designed for image and video object detection, not for audio transcription or general text processing required here.
• D. Stable Diffusion is a generative image model, not suitable for text summarization; Comprehend is not the optimal choice for this specific workflow.

Correct answer: A

Amazon Managed Service for Apache Flink with the built-in RANDOM_CUT_FOREST function provides the lowest operational overhead because it is a fully managed service that requires no infrastructure management, no custom Lambda functions, and no endpoint deployment. The RANDOM_CUT_FOREST algorithm is specifically optimized for real-time anomaly detection on streaming data and scales automatically with Kinesis.

Why the other options are wrong:

• B. While functional, this requires managing SageMaker endpoints and writing custom Lambda code, increasing operational complexity.
• C. Managing Apache Kafka on EC2 instances adds significant operational overhead including instance management, patching, and scaling.
• D. SQS FIFO is designed for message queuing, not real-time streaming; Glue ETL is batch-oriented and introduces latency, making it unsuitable for thousands of JSON records per second.

Correct answer: C

Amazon Comprehend is a fully managed natural language processing service with built-in sentiment analysis capabilities specifically designed for text. It requires no model training or infrastructure management and can immediately analyze chat transcripts to determine sentiment, making it the fastest solution to complete the evaluation.

Why the other options are wrong:

• A. Rekognition analyzes images and videos, not text chat data.
• B. Training a Naive Bayes classifier requires data labeling, feature engineering, and model training, consuming significantly more time.
• D. Random forests require manual feature engineering, data preparation, and model training, adding substantial time overhead.

Correct answer: D

Decreasing temperature reduces randomness in token selection, making responses more deterministic and consistent. Decreasing top_k limits the pool of candidate tokens to the k most likely options, further reducing variability. Together, these changes minimize the randomness that causes similar questions to produce different answers.

Why the other options are wrong:

• A. Increasing temperature increases randomness; increasing top_k expands the token selection pool, both making responses less consistent.
• B. Increasing temperature increases randomness, contrary to the goal of consistency.
• C. While decreasing temperature helps, increasing top_k expands token selection and increases variability, undermining the goal.

Correct answer: C

The target_precision hyperparameter in SageMaker's linear learner allows you to set a precision threshold, which directly minimizes false positives. A higher target_precision value optimizes the decision threshold to prioritize precision (true positives / all positive predictions), reducing incorrect positive weed detections.

Why the other options are wrong:

• A. Setting weight decay to zero affects regularization but does not directly control the precision-recall tradeoff needed to minimize false positives.
• B. Increasing training epochs improves model fit but does not address the classification threshold needed to reduce false positives.
• D. Changing to regressor would change the problem from classification to regression, which is inappropriate for detecting the presence/absence of a weed.

Correct answer: A

Setting Firehose buffer interval to zero eliminates the 60-second batching delay, and tuning the batch size in PutRecordBatch operation allows optimization of throughput while maintaining low latency. This achieves sub-second latency by delivering data to OpenSearch immediately rather than waiting for the buffer interval.

Why the other options are wrong:

• B. DataSync is designed for large-scale data transfer and migration, not real-time streaming; enhanced fan-out is an Kinesis feature unrelated to Firehose.
• C. Increasing the buffer interval to 120 seconds worsens latency, moving in the opposite direction of the requirement.
• D. SQS adds message queue latency and is not designed for sub-second streaming requirements; it would worsen latency.

Correct answer: A

SageMaker real-time inference endpoints provide high availability through multi-AZ deployment, minimum latency through optimized container serving, and built-in auto- scaling that adapts to unpredictable request bursts proportionally. This is a fully managed solution designed exactly for this use case with request sizes between 1 KB and 3 MB.

Why the other options are wrong:

• B. ECS scheduled scaling is time-based, not responsive to unpredictable bursts; it lacks the automatic, demand-driven scaling needed.
• C. Horizontal pod autoscaling based on memory metric is reactive and may not scale quickly enough for unpredictable bursts; EKS adds operational complexity.
• D. Spot Instances are cost-effective but not guaranteed for high availability; manual fleet management adds operational overhead compared to managed SageMaker endpoints.

Correct answer: D

In Amazon EMR, the primary node manages the cluster and must remain stable, so it requires On-Demand Instances. Core nodes store HDFS data and cannot be lost, requiring On-Demand Instances. Task nodes only perform computations and can be interrupted without data loss, making Spot Instances cost-effective for them. This configuration balances cost savings with data durability requirements.

Why the other options are wrong:

• A. Using On-Demand for all instances is unnecessarily expensive when task nodes can use cheaper Spot Instances.
• B. Spot Instances can be terminated, risking loss of the primary node's state and core node HDFS data.
• C. Core nodes store HDFS data and cannot be lost, so they must use On-Demand Instances rather than Spot.

Correct answer: D

SageMaker Pipelines is purpose-built for orchestrating complex multi-step ML workflows with support for diverse data formats, long-running operations, and NLP transformations. It integrates seamlessly with other SageMaker components and can be automated via EventBridge for scheduled or event-driven execution. Data Wrangler (A) has limitations on data size and step complexity. SageMaker notebooks (B) require manual orchestration and don't scale well for automation. Lambda (C) has execution time limits that make it unsuitable for processes taking hours.

Why the other options are wrong:

• A. Data Wrangler has constraints on dataset size and complexity of transformations, and is less suitable for multi-hour processing steps.
• B. Notebooks require manual execution and EventBridge integration is cumbersome; this approach lacks native automation for complex multi-step workflows.
• C. Lambda functions have a 15-minute maximum execution timeout, making them unsuitable for processing that takes hours.

Correct answer: A

AWS::SageMaker::Model defines the ML model artifact and configuration that will be hosted on an endpoint. To host this model, you then declare AWS::SageMaker::Endpoint in the same template, which references the model. The Model resource is the prerequisite that must be declared first to enable endpoint creation.

Why the other options are wrong:

• B. Endpoint resource hosts the model but does not define the model itself; you need Model first.
• C. NotebookInstance is for interactive development and is not used to host trained models.
• D. Pipeline is for orchestrating training workflows, not for hosting deployed models.

Correct answer: C

Lake Formation provides fine-grained access control through tagging mechanisms that map users to resources based on campaign assignments. Lake Formation tags enable column-level and table-level access control in the Glue Data Catalog, which directly controls both Athena query access and S3 bucket visibility. This is the most operationally efficient approach as it centralizes access management in Lake Formation rather than requiring manual policy updates across multiple services.

Why the other options are wrong:

• A. IAM policies alone cannot provide the granular, tag-based access control that Lake Formation offers for campaign-specific data.
• B. Using DynamoDB and Lambda to update S3 bucket policies is operationally complex and lacks the native access control features of Lake Formation.
• D. S3 bucket policies are coarse-grained and difficult to maintain for campaign-specific access patterns; Lake Formation tagging is more scalable.

Correct answer: D

Apache Parquet is a columnar storage format optimized for analytical workloads with built- in compression and efficient data retrieval. It minimizes processing time by allowing SageMaker Canvas to read only the columns needed, skip unnecessary data, and leverage predicate pushdown. For complex, structured data at scale, Parquet significantly outperforms CSV, JSON, and JSONL formats in processing speed and resource efficiency.

Why the other options are wrong:

• A. CSV with Snappy compression is row-based and requires full file parsing, less efficient than columnar Parquet for complex data.
• B. JSONL format is text-based and requires parsing each line individually, slower than Parquet for large datasets.
• C. JSON with gzip compression is text-based and does not provide the columnar optimization benefits of Parquet.

Correct answer: D

Recall measures the proportion of actual positive cases correctly identified (true positives divided by all actual positives). When false negatives are costly-meaning missed positive cases are expensive-high recall is critical to minimize these missed detections. High recall ensures the model catches as many true positive cases as possible, even if it generates more false positives, which are acceptable given the cost structure.

Why the other options are wrong:

• A. Low precision means many false positives; this is not the priority when false negatives are more costly.
• B. High precision reduces false positives, which is less important when false negatives carry higher cost.
• C. Low recall means missing true positives, directly increasing the high-cost false negative rate.

Correct answer: C

CloudTrail records all API calls made to AWS services, including SageMaker endpoint invocations, providing a complete audit trail of endpoint activity. CloudWatch can display this data via dashboards and CloudWatch alarms can trigger notifications when API call counts breach thresholds. This solution directly addresses the requirements for recording API events and threshold-based alerting. SageMaker Debugger (A, B) is designed for training job monitoring, not inference endpoint API call tracking. Option D monitors the built-in Invocations metric but lacks the comprehensive API call recording that CloudTrail provides.

Why the other options are wrong:

• A. SageMaker Debugger tracks training job metrics and tensor values, not endpoint API call events.
• B. SageMaker Debugger and its tensor_variance rule are for training job monitoring, not for recording endpoint invocation API events.
• D. The CloudWatch Invocations metric provides count data but does not record or audit individual API calls like CloudTrail does.

Correct answer: C

SageMaker Pipelines Callback steps are specifically designed to integrate with external services like AWS Glue workflows. A Callback step starts the Glue workflow and pauses the pipeline until the workflow completes, then resumes execution. This requires minimal configuration and no additional orchestration services, making it the least operationally overhead solution compared to managing separate Step Functions orchestration, EventBridge rules, or manually configuring processing steps with ARN inputs.

Why the other options are wrong:

• A. Step Functions would require additional setup and management of separate orchestration logic for both Glue and SageMaker components.
• B. Processing steps in SageMaker Pipelines are designed to run custom code, not to orchestrate external Glue workflows directly.
• D. EventBridge would require creating custom event rules and integration logic, adding unnecessary complexity.

Correct answer: A

Amazon Redshift's dynamic data masking policies provide query-time masking without modifying the source data or creating additional stored copies. This is the most straightforward solution that meets all requirements: sensitive data remains untransformed in the source, no anonymized copies are stored in the database, and access control happens automatically at query time with minimal implementation effort.

Why the other options are wrong:

• B. Materialized views would require storing derived data, and masking logic in views is less flexible than dynamic policies.
• C. Unloading to S3 and using Athena adds complexity and operational overhead compared to native Redshift masking.
• D. Using Glue to anonymize data requires ETL job creation and storing separate anonymized datasets, violating the constraint of not storing anonymized data in or outside the database.

Correct answer: D

SageMaker Debugger includes built-in rules specifically designed to detect vanishing gradients, underutilized GPU, and overfitting problems during training. It provides comprehensive real-time metrics and can be configured to automatically trigger predefined actions when issues are detected, requiring minimal manual setup and no additional Lambda or SNS infrastructure.

Why the other options are wrong:

• A. TensorBoard requires manual monitoring and publishing findings to SNS, then setting up Lambda functions-significantly more operational overhead.
• B. CloudWatch default metrics do not include the deep gradient analysis needed for vanishing gradient detection.
• C. Manually expanding CloudWatch metrics and setting up Lambda automation is more complex than using built-in Debugger rules.

Correct answer: D

Shadow testing deploys a shadow variant alongside the production model, sends live traffic to both models, and compares predictions without affecting production responses. This allows comprehensive performance assessment on real data without any impact on end users, as predictions from the shadow variant are never used to make actual decisions.

Why the other options are wrong:

• A. SageMaker Debugger is for monitoring model training, not for testing deployed model variants in production.
• B. All-at-once traffic shifting would immediately direct traffic from the old model to the new one, affecting end users.
• C. Canary traffic shifting still routes some percentage of live traffic to the new model, risking impact on production users.

Correct answer: C

Partitioning data by date prefix in S3 enables Athena to use partition pruning, dramatically improving query performance for the most recent 3 days of data. This combined with S3 Lifecycle policies to archive older partitions to Glacier provides optimal retrieval performance for the 30-day retention period while managing costs effectively.

Why the other options are wrong:

• A. Unpartitioned data requires Athena to scan millions of rows every time, causing poor performance.
• B. Lambda-based copying and separate buckets adds complexity without the performance benefits of partition pruning.
• D. Creating separate buckets per day is inefficient and prevents Athena from leveraging partition elimination for queries spanning multiple days.

Correct answer: A

For real-time fraud detection, SageMaker real-time inference provides immediate responses to individual transactions, eliminating the delays consumers are experiencing with asynchronous inference. SageMaker Model Monitor provides the necessary notifications about model quality deviations, making this the complete solution for both performance improvement and quality monitoring.

Why the other options are wrong:

• B. Batch transform is designed for processing large datasets in batches, not suitable for real-time transaction fraud detection.
• C. Serverless Inference is appropriate but Inference Recommender provides sizing recommendations, not quality deviation notifications.
• D. Asynchronous Inference is the root cause of the reported delays; switching to it does not solve the performance problem.

Correct answer: D

SageMaker endpoint auto scaling uses CloudWatch metrics to dynamically adjust the number of instances based on demand. This provides automatic scaling during peak usage while allowing instances to scale down to zero during periods of low usage, meeting both cost minimization and capacity requirements.

Why the other options are wrong:

• A. Lambda with fixed concurrency requires pre-allocated capacity, preventing cost savings during low-usage periods.
• B. ECS Fargate with a static number of tasks cannot scale down during low usage, wasting costs.
• C. Manually deploying multiple copies and using an ALB requires manual capacity management and does not automatically scale.

Correct answer: B

AWS Budgets is specifically designed to send automated alerts when spending reaches defined thresholds. Resource tagging should be configured through the SageMaker user profile in the domain, which allows Budgets to track and alert on SageMaker-specific compute costs without requiring IAM profile modifications.

Why the other options are wrong:

• A. Cost Explorer is for analyzing costs after the fact, not for setting threshold-based alerts.
• C. Editing individual IAM profiles is unnecessary and not the standard approach; SageMaker user profile tagging is the proper method.
• D. While Budgets is correct, editing IAM profiles is the wrong configuration method for SageMaker domain-based cost tracking.

Correct answer: D

SageMaker Data Wrangler is purpose-built for data preparation tasks like dropping unnecessary columns with minimal effort. It provides a visual, low-code interface where users can configure transform steps to remove correlated or unwanted columns from Parquet files without writing code. A 50 MB file is well within Data Wrangler's capabilities, making this the least-effort solution.

Why the other options are wrong:

• A. Downloading to a local workstation and writing custom Python scripts requires more manual effort than using Data Wrangler's built-in transforms, and one-hot encoding is not the appropriate technique for dropping columns.
• B. Creating an EMR Spark job requires infrastructure setup and custom scripting, which is considerably more effort than using Data Wrangler's visual interface.
• C. Creating a SageMaker processing job requires writing and managing custom processing scripts, which is more complex than using Data Wrangler's no-code transform steps.

Correct answer: A

Amazon Q Business supports blocked phrases functionality that prevents specific terms from appearing in responses. This is the direct and intended mechanism for excluding competitor names from API responses. Configuring the competitor's name as a blocked phrase ensures it will not be included in any Amazon Q Business responses.

Why the other options are wrong:

• B. Retrievers are designed to control which documents are searched, not to filter specific phrases from responses.
• C. Kendra indexing controls which documents are available for search but does not prevent specific phrases from appearing in final responses.
• D. Document attribute boosting controls document relevance ranking, not phrase-level filtering of response content.

Correct answer: D

SageMaker JumpStart provides pre-trained foundation models with built-in fine-tuning capabilities, and SageMaker Autopilot is designed for automated ML with minimal code. Using Autopilot with LLMs deployed through JumpStart provides a low-code/no-code approach to fine-tuning LLMs for text summarization, requiring minimal manual intervention.

Why the other options are wrong:

• A. SageMaker Studio alone is an IDE and does not provide the low-code/no-code automation that Autopilot offers; EC2 deployments require more manual management.
• B. Custom API endpoints require custom development and infrastructure management, contradicting the low-code/no-code requirement.
• C. EC2 instance management requires significant manual configuration and is not consistent with a low-code/no-code approach.

Correct answer: D

A serverless inference endpoint is ideal for this use case: it runs once per night on a schedule with minimal data (3 MB) and completes in under 1 minute. Serverless endpoints automatically scale to zero when not in use, eliminating costs during idle time. Setting MaxConcurrency to 1 ensures the single nightly prediction request is handled appropriately without unnecessary parallelism.

Why the other options are wrong:

• A. Multi-model serverless endpoints are designed for scenarios with multiple models, not a single model running once daily; caching is unnecessary for a single nightly execution.
• B. Asynchronous inference is designed for high-volume, latency-tolerant workloads; it incurs costs for the infrastructure even when idle, and InitialInstanceCount parameter doesn't apply to async endpoints.
• C. Real-time endpoints maintain instances continuously, incurring unnecessary costs when the model is idle for 23+ hours per day; auto-scaling to 0 defeats the purpose of a real-time endpoint.