Content

Target Audience: Large to mid-sized internet companies, AI companies, and research or academic institutions. New Feature/Specification: Autoscaling is now supported for non-Head roles in Ray on DLC. Users can set a minimum (min) and maximum (max) number of instances. The min value ensures a fast job start, after which the system will attempt to scale out to the max instance count. This feature integrates with quota preemption. When a task is preempted by a higher-priority one (from a parent or high-priority peer quota), the scheduler will first deallocate the scaled-out role instances. This approach keeps more tasks running concurrently, optimizing for both system-wide job throughput and resource utilization.