Panel Discussion: Fragmentation of the Batch Ecosystem in Kubernetes, Challenges and Solutions
summary
Panelists compare concrete systems and limits: Armada, Unicorn, MCAD and Flux (qflux/fluence) build schedulers, queuing, quotas and hierarchical borrowing because etcd chokes at ~100k simultaneous job submissions and Kubernetes events lack ordering or delivery guarantees. Armada reports scaling two orders of magnitude beyond etcd by buffering, while the Flux operator instantiates a mini cluster inside Kubernetes to run tens‑to‑hundreds of thousands of tasks. Remaining gaps: gang scheduling fragmentation, multi‑cluster federation, multi‑tenant quotas and shared filesystem health.
AI-written summary · about
chapters (22)
- 0:00Batch workloads on Kubernetes
- 2:06Declarative vs imperative batch models
- 5:47Challenges in batch workloads
- 7:51Missing batch features vs YARN
- 10:00Kubernetes for batch workloads?
- 13:52Solutions for batch scheduling gaps
- 15:34Batch vs Kubernetes adaptation
- 18:11Power user needs vs batch efficiency
- 21:01Multi-cluster batch federation
- 25:34Mathematical accuracy in batch workloads
- 27:02UI and identity for batch systems
- 31:54Multi-tenancy in batch Kubernetes
- 34:01Scaling etcd and dependencies
- 36:26Collaboration with traditional schedulers
- 38:09Improving etcd for batch frameworks
- 41:49Educating users for batch submission
- 43:15User workload correlation in Kubernetes
- 45:22Storage challenges for batch workloads
- 46:24Defending etcd and tools
- 47:50Common APIs for batch workloads
- 50:18Forums for batch Kubernetes discussions
- 51:05Closing remarks
related talks
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
33mBatch Systems in Production with Kueue: Multi-Tenancy and Fungibility
36mManaging Thousands of Clusters and Their Workloads with Flux
42mMastering Kubernetes: Challenges, Strategies, and Best Practices
39mPanel: Cracking the Data Locality Puzzle
34mPanel: Improving and Managing Kubernetes at Scale
01h 06mPanel Discussion: Cloud Foundry Test Kitchen
33mPanel Sessions with Keynote Speakers
