Resource Fairness and Utilization for Heterogeneous Batch/ML Platforms With Kueue

Explore resource fairness and utilization challenges in heterogeneous batch and machine learning platforms through this 10-minute lightning talk from KubeCon + CloudNativeCon. Learn how to achieve high resource utilization while maintaining fairness across different teams and workload types using Kueue, a Kubernetes-native job scheduler. Discover the limitations of Kubernetes' priority-based preemption at the pod level and understand how timing issues can create unfair resource allocations when early-submitting teams monopolize resources. Examine two key Kueue features: Hierarchical Fair Sharing and History-Based Fair Sharing, including the algorithms that power these capabilities and optimizations designed for high-throughput scheduling environments. Gain insights into addressing the critical balance between preventing resource waste and ensuring equitable access for different tenant types, including those requiring immediate resource access and varying workload categories such as training versus inference operations.