An ML specialist is preparing data with a PySpark job that includes complex window aggregations. They want to test how different feature counts and sample sizes affect model performance. Which approach lets them record transformation parameters and results so they can evaluate and compare data preparation variants?

The correct answer is to use an Amazon SageMaker Experiments tracker in the script to record key parameters and run the script as a SageMaker processing job. SageMaker Experiments is designed for tracking, organizing, and comparing machine learning experiments, including data preparation steps. It allows you to log parameters (like feature counts and sample sizes) and metrics, making it ideal for evaluating different data preparation variants. A SageMaker processing job is suitable for running PySpark scripts with complex window aggregations. Incorrect options: SageMaker Debugger is primarily for debugging training jobs by capturing tensors and profiling system resource utilization, not for tracking data preparation parameters and results. While AWS Glue jobs can run PySpark, SageMaker Experiments integrates seamlessly with SageMaker processing jobs for comprehensive experiment tracking within the SageMaker ecosystem.