Kubernetes Workload Resiliency in Action: Beyond Basics
summary
Shows how specific Kubernetes levers combine to keep long AI jobs and multi-tenant clusters alive: PDBs turn hard kills into graceful drains with preStop hooks and checkpoints; use priority classes for eviction order (not scheduling) and QoS tiering to avoid surprising preemptions; resource requests/limits, resource quotas, runtime classes, affinity, DRA, gang scheduling and topology manager are demonstrated; cites 66% vs 7% (can run AI vs deploy daily) and a 512‑GPU example where 400 bound GPUs cause massive waste; warns of NUMA slowdowns, rounding bugs at scale, and cascading failures unless node allocatable and system reservations are profiled.
AI-written summary · about
chapters (9)
- 0:00Kubernetes workload resiliency intro
- 3:15Why resiliency matters beyond basics
- 8:08Training workloads under pressure
- 11:40Reliable inference services
- 15:53Multi-tenancy challenges
- 19:44Kubernetes levers in practice
- 21:55Resiliency breaks at scale
- 26:27Five key takeaways
- 28:27Q&A
related talks
45mKubernetes Resiliency
38mKubernetes for AI Workloads – What Works and What… Doesn’t
59mKubernetes - going beyond the basics
01h 01mKubernetes – going beyond the basics
16mAI + Kubernetes: What Beginners...
37mCo-Location of CPU and GPU Workloads with High Resource Efficiency
09mKeynote: Kubernetes for GenAI: The New Operating System of the AI-First World
27mHarnessing Kubernetes for Scalable AI/ML Workloads