SIG Scheduling Update: Transition From Pod To Workload Scheduling
summary
Kubernetes is shifting from pod-centric to Workload-aware scheduling with new Workload and PodGroup APIs, a PodGroup scheduling cycle and PodGroup meanCount semantics that guarantee at-least-N bindings (example: bind 2 of 3 pods, or none). Preemption is reimplemented for workloads with a disruptionMode that can preempt single pods or entire groups. Topology-aware placements evaluate racks/zones per-placement. Kueue integration faces decision drift (real example: 16 GPUs cluster, two nodes with 8 free, two 4-GPU pods — Q admits and reserves 8 GPUs but kube-scheduler rejects due to missing anti-affinity), so a schedule library phase aligns Q’s simulation with kube-scheduler predicates and exposes prefilter/filter/reserve extension points.
AI-written summary · about
chapters (8)
- 0:00SIG Scheduling Update intro
- 3:27Kube Scheduling overview
- 8:51Workload-aware scheduling
- 13:51Preemption and topology
- 17:55PodGroup and advanced scheduling
- 21:08Custom scheduler challenges
- 25:53Schedule Library consolidation
- 30:59Q&A
related talks
30mSIG Scheduling Intro & Updates
30mDeep Dive: Scheduling SIG
35mSIG Scheduling Deep Dive
10mFair Scheduling for Deep Learning Workloads in Kubernetes
34mKubernetes Scheduling Features or How Can I Make the System Do What I Want?
36mDeep Dive: SIG Scheduling
29mSIG Scheduling Deep Dive
42mDeep Dive Into the Latest Kubernetes Scheduler Features