A. Incorrect.
EMR and notebooks require technical coding or Spark knowledge.

B. Correct.
SageMaker Canvas with Data Wrangler provides visual, low-code data preparation for tabular ML data.

C. Incorrect.
SQL and pandas require coding skills, which the team does not have.

D. Incorrect.
A notebook workflow still requires programming expertise.

A. Correct.
Athena is serverless and can create/query partitioned tables over S3 data with minimal infrastructure.

B. Incorrect.
Replication and Object Lambda add unnecessary moving parts and do not provide SQL analytics like Athena.

C. Incorrect.
Glue Spark can work, but it requires job management and is more overhead for ad hoc date-based queries.

D. Incorrect.
Firehose is for streaming delivery, not querying existing S3 CSV objects.

A. Incorrect.
Dropping rows reduces data completeness and can discard useful examples.

B. Incorrect.
The mean is sensitive to extreme outliers and can distort imputed values.

C. Correct.
Median imputation preserves rows and columns while being more robust to outliers.

D. Incorrect.
Dropping columns removes potentially important features and does not maximize completeness.

A. Incorrect.
Glue can run data jobs, but it is not the managed SageMaker inference and model-quality monitoring path.

B. Incorrect.
CloudTrail audits API activity; it does not monitor model data quality.

C. Incorrect.
ECS would require custom ML inference and monitoring logic.

D. Correct.
Batch transform supports offline inference, and Model Monitor can schedule data-quality checks and alert through CloudWatch.

A. Incorrect.
This creates a custom pipeline and SQL storage, but SQL over Glue metadata is not semantic search.

B. Incorrect.
Feature Store is not a semantic search engine or vector database for document retrieval.

C. Correct.
Amazon Kendra provides managed semantic search and can ingest documents directly from S3.

D. Incorrect.
Textract extracts text from documents and images; it does not provide semantic search.

A. Incorrect.
A multi-model endpoint is for hosting many models, not primarily for one small nightly invocation.

B. Incorrect.
Asynchronous inference is useful for larger payloads or longer processing, but this small, short job fits serverless inference.

C. Incorrect.
Real-time endpoints generally keep provisioned capacity and are not the most cost-efficient for one nightly request.

D. Correct.
Serverless inference avoids idle capacity and MaxConcurrency of 1 fits a single scheduled invocation.

A. Incorrect.
Batch transform is better for offline batch jobs and does not directly fit per-request completion notifications.

B. Correct.
SageMaker asynchronous inference supports large inputs in S3 and can publish completion notifications to SNS.

C. Incorrect.
EFS is unnecessary because SageMaker async inference integrates with S3 for input and output.

D. Incorrect.
The inference pattern is right, but EFS is not needed and S3 is the native input/output store.

A. Incorrect.
Glue is for cataloging and ETL, not native ML model drift detection.

B. Incorrect.
Flink adds unnecessary streaming infrastructure for a managed SageMaker monitoring problem.

C. Correct.
Model Monitor detects drift and can emit events that Lambda uses to start a SageMaker pipeline.

D. Incorrect.
QuickSight anomaly detection is for BI analytics, not production ML model drift monitoring.

A. Incorrect.
SSH keys are for authentication, not centralized data encryption and key access control.

B. Incorrect.
API keys are not encryption keys.

C. Incorrect.
An IAM role can grant permissions, but it does not create or manage the encryption key by itself.

D. Correct.
A customer managed KMS key provides centralized encryption key management, policy control, and auditability.

A. Incorrect.
Firehose is primarily for streaming delivery and is not the broad data integration choice for S3-based ingestion pipelines.

B. Correct.
AWS Glue provides serverless data ingestion and transformation, and SageMaker Studio Classic supports ML build and deployment workflows.

C. Incorrect.
Redshift ML is for building models from Redshift data, not creating general ingestion pipelines from S3.

D. Incorrect.
Athena queries data but does not create ingestion pipelines, and notebooks are less suitable as deployment pipelines.