A data engineering team runs a nightly AWS Glue ETL job that cleans and joins several large CSV files from Amazon S3. The transformed output will be queried repeatedly by data scientists using Amazon Athena to select a small subset of columns for feature exploration, and later read by a SageMaker training job. The team wants to minimize both Athena query cost and downstream read time. In what format should the Glue job write the transformed data to S3?