You have IoT soil sensor data in a 10 GB Amazon DynamoDB table and weather event data as 5 GB of JSON files in Amazon S3. You want to prepare this combined dataset to train an Amazon SageMaker model with the least administrative overhead. Which approach best accomplishes the required transformation?

The correct approach leverages AWS Glue for its serverless nature and integrated data cataloging, minimizing administrative overhead. AWS Glue crawlers automatically infer schemas and catalog both the DynamoDB table and S3 JSON files, making them queryable. An AWS Glue ETL job (using Apache Spark) can then efficiently merge these datasets. Writing the output to S3 as CSV files is a common and flexible format for SageMaker model training. Starting an Amazon EMR cluster involves managing servers and is more administrative overhead than Glue. Writing to Amazon Redshift is unnecessary for SageMaker training data, which typically resides in S3. Using DynamoDB Streams and Lambda to append to existing S3 files is inefficient for batch processing and merging, and could lead to data inconsistencies or corruption with concurrent writes.