Files
kubernetes/cluster/addons
Kubernetes Submit Queue 1e76d9e1d5 Merge pull request #47356 from dashpole/master_critical_pods
Automatic merge from submit-queue (batch tested with PRs 47669, 40284, 47356, 47458, 47701)

Mark Static pods on the Master as critical

fixes #47277.

A known issue with static pods is that they do not interact well with evictions.  If a static pod is evicted or oom killed, then it will never be recreated.  To mitigate this, we do not evict static pods that are critical.  In addition, non-critical pods are candidates for preemption if a critical pod is scheduled to the node.  If there are not enough allocatable resources on the node, this causes the static pod to be preempted.

This PR marks all static pods in the kube-system namspace as critical.

cc @vishh @dchen1107
2017-06-19 15:25:01 -07:00
..
2017-05-11 12:29:00 -07:00

Cluster add-ons

Overview

Cluster add-ons are resources like Services and Deployments (with pods) that are shipped with the Kubernetes binaries and are considered an inherent part of the Kubernetes clusters.

There are currently two classes of add-ons:

  • Add-ons that will be reconciled.
  • Add-ons that will be created if they don't exist.

More details could be found in addon-manager/README.md.

Cooperating Horizontal / Vertical Auto-Scaling with "reconcile class addons"

"Reconcile" class addons will be periodically reconciled to the original state given by the initial config. In order to make Horizontal / Vertical Auto-scaling functional, the related fields in config should be left unset. More specifically, leave replicas in ReplicationController / Deployment / ReplicaSet unset for Horizontal Scaling, leave resources for container unset for Vertical Scaling. The periodic reconcile won't clobbered these fields, hence they could be managed by Horizontal / Vertical Auto-scaler.

Add-on naming

The suggested naming for most of the resources is <basename> (with no version number). Though resources like Pod, ReplicationController and DaemonSet are exceptional. It would be hard to update Pod because many fields in Pod are immutable. For ReplicationController and DaemonSet, in-place update may not trigger the underlying pods to be re-created. You probably need to change their names during update to trigger a complete deletion and creation.

Analytics