WG-Batch Updates: What’s New and What Is Next?
summary
Kubernetes batch work focuses on concrete changes: Job API reaches suspended jobs beta in 1.36, work creation and PodGroup auto-creation appear in 1.37 alpha, and 1.38 plans fine‑grained job failure reasons. JobSet adds elastic job set and an impress restart protocol that shrinks recovery from minutes to seconds for 1k–100k‑pod trainings by restarting containers only. Kueue (2,700 stars, ~200 contributors) demos Multi-Cluster dispatching, fair sharing, topology‑aware scheduling, orchestrated preemptions, concurrent admission and scheduler‑library integration with SIG Scheduling.
AI-written summary · about
chapters (13)
- 0:00WG-Batch Updates intro
- 0:21Job API evolution in Kubernetes
- 1:22Job Set for HPC and AI
- 2:04Impress restart feature
- 4:57Kubernetes Scheduler overview
- 6:45Device integration in Q
- 10:54MultiQ for scaling
- 12:42Topology-aware scheduling in Q
- 14:48Concurrent admission in Q
- 16:34Recent Q features and plans
- 18:20Workload API integration
- 23:40Q and Workload API
- 27:33Community Q&A
related talks
34mWG-Batch Updates: What’s New and What Is Next?
37mIntroduction to the Kubernetes WG Batch
16mKubernetes SIG Apps Updates
27mKubernetes SIG Apps Updates
29mLightning Talk: Evolving the Job API to Become Defacto Standard for Batch Worklo...
05mClosing Remarks
26mKubernetes SIG Apps Updates
30mSIG Scheduling Intro & Updates