FinPredict runs many consecutive training jobs with the same instance type and would like to reduce startup latency. They consider enabling SageMaker Warm Pools to reuse provisioned infrastructure across jobs. Which statement correctly describes what warm pools provide and an important limitation to plan for?

The correct answer accurately describes SageMaker Warm Pools. Warm pools reduce startup latency by keeping EC2 instances provisioned for a configurable duration after a job completes, making them available for subsequent jobs of the same instance type. This eliminates the time needed for instance provisioning. However, warm pools do not persist the job's file system (like /opt/ml) across jobs. Data, checkpoints, or models needed for subsequent jobs must still be saved to external persistent storage like Amazon S3 or Amazon EBS. The first incorrect option is wrong because warm pools do not automatically persist the container filesystem; this requires external storage. The second incorrect option is wrong because warm pools are tied to specific instance types and Availability Zones; they cannot be used interchangeably across different types or zones. The third incorrect option is wrong because warm pools are available for both on-demand and managed spot training jobs and can be configured with a keep-alive duration.