A company stores training data in nested JSON files in S3 and needs a tabular format for XGBoost training. Which approach has the least operational overhead to convert those JSON files to tabular data?

AWS Glue PySpark jobs with the Relationalize transform are designed specifically for converting nested semi-structured data like JSON into a flat, tabular format suitable for machine learning. This approach minimizes operational overhead because Glue is a fully managed, serverless ETL service, and the Relationalize transform automates much of the complex schema flattening logic. Writing custom Scala code for Amazon EMR Serverless would involve more development effort and managing EMR configurations, even if serverless. An AWS Lambda function using reduce() for large files would likely hit memory or execution time limits and is not optimized for large-scale data transformations. While Amazon Athena can query JSON, its flatten function is primarily for arrays within a single column, and creating a robust tabular structure from deeply nested JSON across many files would be complex and less efficient than Glue's dedicated transform.