A company extracts about 1 TB of data per day from sources such as SAP HANA, SQL Server, MongoDB, Kafka, and DynamoDB. Some sources have undefined or evolving schemas. The solution must detect schemas, perform ETL, and load the data into S3 within 15 minutes of data creation. Which approach provides the required functionality with the least operational overhead?

AWS Glue is the most suitable choice because it is a serverless ETL service that natively supports schema detection and evolution, which is crucial for sources with undefined or evolving schemas. It integrates with various data sources, including SAP HANA, SQL Server, MongoDB, Kafka, and DynamoDB, and can process 1 TB of data daily within the 15-minute latency requirement. Creating an Apache Spark pipeline within Glue leverages its managed Spark environment, minimizing operational overhead compared to self-managing EMR. Using Amazon EMR would require managing the EMR cluster, increasing operational overhead. Creating a PySpark program in AWS Lambda is not suitable for 1 TB of daily data processing due to Lambda's execution duration and memory limits. Creating a stored procedure in Amazon Redshift is designed for data already in Redshift and not for initial ingestion and ETL from diverse external sources into S3.