Stick Shift or Automatic? How to Drive a Data Center Operating System
summary
Argues that Apache Mesos turns a datacenter into one big computer by outsourcing placement to frameworks like Aurora and Marathon, giving dynamic placement for free. Lists concrete Mesos features: unified containerizer for Docker/rkt images, event stream and operator API, task unreachable handling (1.1) to signal partitions, maintenance primitives (0.25) to drain nodes, quotas/reservations, persistent volumes for Cassandra, and remote attach/exec for debugging. Shows these reduce manual owner-tracking and monitoring pain, while surfacing new issues like resource heterogeneity and container sizing.
AI-written summary · about
chapters (30)
- 0:00Intro
- 0:35Challenges
- 1:31Data Center Operator
- 2:27Iterating
- 3:28Indirection
- 3:58Methods Master
- 4:56Schedulers
- 5:24No Machine is Special
- 6:02Normal with Missiles
- 6:49Problems
- 7:48Deployment
- 8:20Cluster Management
- 9:03Dynamic Placement
- 9:43Dependencies
- 11:08Monitoring
- 12:40Tasks
- 13:16Monitor
- 13:51Events
- 15:03Problem Logging
- 16:15Custom Logging Module
- 19:33Other Source Homogeneity
- 20:54Containers
- 21:34Mindset Change
- 21:52Failure
- 23:56Network failures
- 24:41Network partition policy
- 25:10No way to determine when a task is not running
- 26:48Time in maintenance
- 27:15Maintenance of limited
- 28:52Capacity planning