Intro: Scheduling SIG
Talkat KubeCon + CloudNativeCon North America 2019 · on Wednesday, 2019-11-20· at 14:25· Room 7AB - San Diego Convention Center, Upper Level·
summary
Explains the scheduler’s actual workflow: informer-based cache, priority queue, optimistic binding with asynchronous HTTP bind calls, and filter+score phases; details that the scheduling framework (1.15/1.16) refactors predicates/priorities into in-process plugins (replacing slow HTTP extenders), and that preemption is GA since 1.14. Covers descheduler eviction strategies (duplicate pods, thresholds, PDB-safe evictions), batch/Co-scheduling projects (kube-batch, Poseidon), and measured performance like 100 pods/sec throughput and a 20x pod anti-affinity improvement.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:25Scope of Scheduler
- 3:54Detailed Scheduling Flow
- 8:38Phase 1: Predicates (a.k.a Filtering)
- 10:09Phase 2: Priorities (a.k.a. Scoring)
- 11:31Recent Developments in default scheduler
- 15:07Design Rationale of Scheduler
- 18:54Descheduler. Why we need it?
- 20:42Descheduler strategies
- 21:38Pod Eviction Restriction
- 22:17Overview of kube-batch
- 23:16Features of kube-batch
- 25:00Features of Poseidon
