Evolving Kubernetes Scheduling

Explore the evolution of Kubernetes scheduling in this 34-minute conference talk that examines how modern AI workloads are driving changes to container orchestration. Learn about the current limitations of Kubernetes' core scheduler and discover how second-level schedulers like Kueue, Volcano, Ray, and Slurm are filling critical gaps in workload-aware scheduling, preemption capabilities, and topology-aware resource allocation. Understand the complexities these extensions face in terms of development, maintenance, and interoperability challenges. Gain insights into comparative analysis of different second-level scheduling solutions and their integration patterns with kube-scheduler. Discover upcoming enhancements to the core Kubernetes scheduler including resource reservations, workload-awareness features, and improved integration with infrastructure autoscaling systems. Examine how these architectural changes will establish clearer separation of responsibilities between core and second-level schedulers, ultimately enabling better utilization of expensive accelerator hardware while ensuring fair resource sharing across teams and supporting complex parallel training and inference workloads.