Batch Systems in Production with Kueue: Multi-Tenancy and Fungibility
summary
Kueue is a Kubernetes-native queueing layer that builds multi-tenant batch clusters by enforcing quotas with borrowing and preemption and mapping workloads to ordered resource flavors (reservation → on‑demand → spot). It runs on‑prem and in autoscaled cloud setups, integrates with Cluster Autoscaler via provisioning requests, supports Job/JobSet/RayJob and gang jobs (launcher+workers), and is deployed at scale (CyberAgent reports a heterogeneous GPU cluster with 200+ namespaces and H100/A100 hardware). Recent release: 0.5 with extensible CRD adapters.
AI-written summary · about
chapters (13)
- 0:00Batch systems with Kueue
- 1:15Job queuing for training workloads
- 2:25Kueue resource quotas
- 3:30Fungibility in Kueue
- 4:35Kueue job APIs and extensibility
- 6:03Kueue as Kubernetes-native
- 7:51Kueue 0.4 vs 0.5 differences
- 12:09CyberAgent’s Kueue use case
- 17:07Capacity management in Kueue
- 18:59Cloud provider reference architecture
- 22:33Fungibility and preemption in queues
- 27:36Kueue future plans
- 30:23Q&A: Kueue vs traditional schedulers
related talks
29mBuilding a Batch System for the Cloud with Kueue
28mMultitenancy and Fairness at Scale with Kueue: A Case Study
35mKueue: A Kubernetes-native Job Queueing
34mWG-Batch Updates: What’s New and What Is Next?
24mAccelerate Your AI/ML Workloads With Topology-Aware Scheduling in Kueue
24mVolcano – Cloud Native Batch System for AI, BigData and HPC
24mKubernetes Batch Processing at Scale - A Scheduling Perspective
37mIntroduction to the Kubernetes WG Batch