Extra 10: Cluster Management with Borg and Kubernetes

  • Borg

  • Google Data Centers

  • Even Bigger

  • program to run
  • cell/custesr
  • cmdline args

  • 10,000 copies of hello world

  • Borg
  • BorgMaster
  • Borglet
  • all managed by Config file
  • And Blaze?

  • things can go wrong

  • a cell
    • usage - cpu/vm

  • lots of waste
  • overhead

  • bigger cells are better, less partitioning overhead

  • bucket to next order of 2

  • some people run very little resource
  • some people have nice round numbers
  • used best fit scheduling algorithm
    • but it could be gamed

  • people overallocate

  • be more aggressive

  • yellow safety margin
    • smaller on week 2
    • then set it back to original settings

  • configs
  • monitor
  • binaries
  • you only need to focus on the app

Kubernetes

  • open source Borg

  • it schedules and allocates
    • you don't need to car about the machine

  • replication

  • Google Kubernetes Service

  • resiliency
  • efficiency
  • containers