Notes on master instance sizing

Allocations of CPU on the master

Note these are only requests, not limits.

  1. 50m  dns-controller
  2. 200m etcd main
  3. 100m etcd events
  4. 150m kube-apiserver
  5. 100m kube-controller-manager
  6. 100m kube-proxy
  7. 100m kube-scheduler
  9. ====
  11. 800m total

  • One a 1 core master, this leaves 200m for misc services e.g. CNI controller, log infrastructure etc. That will be less if we start reserving capacity on the master.

  • kube-dns is relatively CPU hungry, and runs on the nodes.

  • We restrict CNI controllers to 100m. If a controller needs more, it can support a user-settable option.

  • Setting a resource limit is a bad idea: https://github.com/kubernetes/kubernetes/issues/51135