Kueue
Kubernetes batch job queue system. CNCF project under sig-scheduling, GA in 2024. Addresses that the Kubernetes default scheduler is pod-oriented, not job-oriented — lacks the fairness, quota, gang scheduling that HPC/ML need.
Concepts: ClusterQueue (quota pool), LocalQueue (namespace-bound), Workload (gang of pods). Cohort feature: multiple ClusterQueues can share resources with priority + preemption. Used heavily for ML training: 8 GPUs at once (gang) or nothing — run a Workload via Kueue instead of letting the default scheduler partial-schedule. Integrates with JobSet, Argo Workflows, Ray. Competitors: Volcano (CNCF incubating), Yunikorn (Apache).