Skip to main content
Quick reference
  • drain = terminate pods + reschedule on other nodes + cordon the node
  • cordon = mark node as unschedulable only

​
Introduction

Assume you have a cluster with 2 nodes and you want to upgrade one of the nodes to a new version. If the node is down for more than 5 minutes, all pods running on that node will be terminated. If the node comes back within that window, pods will come back online when the kubelet process starts. Pods that are part of a ReplicaSet will be provisioned on other nodes automatically. To upgrade a node safely, you must drain it first.

​
Draining a node

When you drain a node, pods are gracefully terminated and rescheduled on other nodes. The node is also marked as unschedulable (cordoned), so no new pods are scheduled on it until you uncordon it.
Draining a node does not move pods — it terminates them and lets the scheduler place new ones elsewhere. Pods not managed by a controller (e.g., bare pods) will be lost unless you use --force.
# Drain a node (gracefully evicts pods managed by a controller)
kubectl drain <node-name>

# Force drain — bare pods will be lost permanently
kubectl drain <node-name> --force

# Ignore DaemonSet-managed pods during drain
kubectl drain <node-name> --ignore-daemonsets

​
Uncordoning a node

After the node comes back online, uncordon it so that new pods can be scheduled on it again. 
kubectl uncordon <node-name>
Pods that were rescheduled to other nodes during the drain will not automatically move back to the original node after uncordoning. They remain on whichever nodes the scheduler placed them.

​
Cordoning a node without draining

If you only want to prevent new pods from being scheduled on a node without evicting existing ones, use cordon. 
kubectl cordon <node-name>
This marks the node as unschedulable without terminating any running pods.

Build docs developers (and LLMs) love

Get started for freeTalk to us