Systems @Scale Summer Q&A Week 2
summary
Kubernetes is evaluated alongside Mesos and Yarn with concrete tradeoffs: GKE tops at ~15,000 nodes but operators prefer many 1,000 node clusters for fault isolation, AKS shares a free control plane by pooling infrastructure, and Optimus offers elastic capacity with evict batch in <5 minutes and SLAs up to the 90s for higher tiers. Facebook uses power‑efficient small servers plus Dynamo for real‑time power capping, Uber experiments with overcommit 10–20%, and apps must handle preemption and graceful shutdown.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:48Class management systems future
- 5:10Dynamic resource sharing
- 9:04AKS resource sharing
- 11:33Workload performance measurement
- 15:27Sudden workload spikes
- 18:15AKS workload troubleshooting
- 21:02Preemption strategies
- 24:16Optimus vs spot instances
- 25:19Capacity planning ratios
- 27:06Kubernetes and Azure
- 29:11Microservices auto scaling
- 30:14Power constraints in leasing
- 32:25Event-driven capacity adjustments
- 33:27VMs vs containers
- 34:53Buffer allocation
- 36:20AKS extensibility
- 38:10Elastic capacity confidence
- 39:18Kubernetes vs Mesos
- 42:06Conclusion
related talks
27mKeynote Panel: Future of Cluster Management
43mScalability Intro – Shyam Jeedigunta, Google & Bob Wise, AWS (Any Skill Level)
34mDistributed applications and Kubernetes
50mScale Kubernetes to Manage 60K Nodes Through Architectural Extension
33mDistributed Scheduler Hell
31mManaging Multi-Cloud Apache Spark on Kubernetes
36mUnderstanding Scalability and Performance in the Kubernetes Master
39mBuilding Distributed Systems with Kubernetes