Leveraging the Kubernetes Enqueue Scheduling Plugin for Smarter Workload Placement
summary
Describes Outrigger as a framework that uses Kubernetes’ scheduling gates via the Enqueue plugin to let controllers asynchronously patch pod specs before scheduling. Uses the multiarch tuning operator and Q on a mixed AMD64/ARM64 cluster with a cluster-autoscaler set to a 2:1 ARM preference. Prototype works when the operator injects nodeAffinity before Q or autoscaler act, but behavior becomes random when Q reconciles higher-level jobs first, exposing a race that breaks autoscaling and flavor assignment.
AI-written summary · about
chapters (9)
- 0:00Kubernetes scheduling challenges
- 3:37Scheduling framework architecture
- 10:31Operational research in scheduling
- 15:23Scheduling gates and orchestration
- 21:11Outtrigger project for smarter scheduling
- 22:56Case study with Q and multiarching
- 28:40Multi-controller scheduling challenges
- 30:29Observability and future of scheduling
- 31:53Q&A on custom schedulers
related talks
32mCoordinate Workloads Colocation: QoS-Oriented Scheduling Enhancement on K8s
28mMultitenancy and Fairness at Scale with Kueue: A Case Study
31mSIG Scheduling Update: Transition From Pod To Workload Scheduling
30mEnhancing the Kubernetes Scheduler for Diverse Workloads in Large Clusters
29mScale Your Batch / Big Data / AI Workloads Beyond the K8s...
29mKubernetes Workload Resiliency in Action: Beyond Basics
24mAccelerate Your AI/ML Workloads With Topology-Aware Scheduling in Kueue
36mNetwork-aware Scheduling in Kubernetes