A geospatial company stores daily satellite images in Amazon S3 for vessel detection. Training data grows incrementally and a SageMaker training job using a single ml.p2.xlarge instance in File input mode recently began failing due to insufficient local storage; the only change was more data. The solution must fix the storage issue while optimizing performance and minimizing training cost. What change should the ML specialist make?

The correct answer is to change the training configuration to use Pipe input mode. Pipe mode streams data directly to the training instance as it's needed, avoiding the need to download the entire dataset to local storage before training begins. This directly addresses the "insufficient local storage" issue, as only a small buffer of data is stored locally at any given time. It also optimizes performance by starting training faster and minimizes cost by not requiring larger, more expensive instances solely for storage capacity. Using two ml.p2.xlarge instances would increase compute but not solve the local storage problem if each instance still tries to download the full dataset. A single ml.p3.2xlarge instance offers more compute and memory but doesn't necessarily increase local disk space enough to handle a continuously growing dataset, and it would be more expensive. Using Amazon EFS would provide shared storage, but it's generally slower than S3 for large-scale data transfer to SageMaker training jobs and would introduce additional cost and complexity without necessarily outperforming Pipe mode for this specific use case.