A Hybrid Container Cloud With Kubernetes and Hadoop YARN
summary
Alibaba builds a colocation framework that runs Kubernetes and Hadoop YARN on one cluster, using a control plane that issues commands to per-node agents, enforces cgroups, and dynamically resizes node allocations so online services keep priority while batch jobs fill idle capacity. The system reports elastic resource sharing and raises average CPU use from 10% to 40%, keeps tail online latency impact under RT impact within 5%, supports YARN queue priorities and noncontainer resource localization, and handles promotion-day autoscaling successfully.
AI-written summary · about
chapters (17)
- 0:00Intro
- 1:15Data center AVG utilization
- 2:23Workloads in Alibaba data center
- 3:39Before: service & Job separated cluster
- 4:27Services vs Jobs
- 7:20Google Borg
- 8:20What is co-location
- 16:08resource sharing
- 17:18CPU ogroup
- 19:03Together time sliced scheduling
- 19:48Normal day Sales day
- 22:33Design Principal
- 24:06API Design
- 24:49Scheduling Unit
- 26:20Communication
- 27:38Container Runtime
- 28:52Larger scale
related talks
20mA full-scenario colocation of workloads based on Kubernetes
22mKeynote: Cloud Native at Alibaba Scale
44mSome Lessons We Learned from Moving E-business Giant to Cloud Native
24mDynamic Big Data in the Cloud with Kubernetes
41mDocker to kubernetes: running Hadoop in a cloud native way
33mKubernetizing Big Data and ML Workloads at Uber
23mPractice of Fine-grained Cgroups Resources Scheduling in Kubernetes
08mLightning Talk: Orchestrating Kubernetes Clusters on HPC Infrastructure