Nodes

A node's capacity, conditions and taints, a root shell on the node, and cordon, uncordon and drain.

Nodes is under Cluster in the sidebar. The table shows each node’s status, instance type, zone, Karpenter NodePool and capacity type, pod count, capacity, Kubernetes version, and CPU and memory use when metrics-server is installed. Filter with the column keys: capacity-type:spot finds spot nodes, and zone:eu-west-1a finds one zone’s.

A node’s drawer

  • Summary: its OS, architecture, kernel, container runtime, kubelet version, pod CIDR, provider ID, zone and region; Capacity / Allocatable for CPU, memory and pods; Conditions (Ready, MemoryPressure, DiskPressure, PIDPressure); and Taints.
  • Terminal: a root shell on the node itself. See below.
  • YAML: the live object.

Node shell

Open a node’s Terminal tab and click Start node shell.

Kubebay starts a short-lived helper pod pinned to that node. The pod is privileged, uses the host’s PID, IPC and network namespaces, and tolerates every taint. It runs nsenter -t 1 -m -u -i -n -p to enter the node’s own namespaces. That gives you a real root shell on the node: its filesystem, processes and network, not the helper pod’s. It’s the same approach Freelens takes.

The helper pod is deleted when you close the drawer. If Kubebay or your browser stops before that, the pod stops itself: it carries a one-hour deadline, after which the cluster ends it, so a privileged pod is never left running. A shell session ends at that deadline too; start another to continue.

The helper pod runs in the default namespace. Its image is registry.k8s.io/e2e-test-images/busybox:1.29-2 unless you change it in Settings, for example to a mirror your cluster can pull from. KUBEBAY_NODE_SHELL_IMAGE sets it for an engine without Settings, such as one in a cluster. The image needs nsenter.

A node shell is root on the node. Your identity needs permission to create privileged pods in default, and your cluster’s admission policies, such as Pod Security restricted or baseline, may refuse them.

Cordon, uncordon and drain

  • Cordon marks the node unschedulable, so no new pods land on it. Uncordon undoes that.
  • Drain cordons the node, then evicts its pods through the Eviction API. DaemonSet pods, static (mirror) pods and finished pods are skipped. Evictions respect PodDisruptionBudgets: if a PDB blocks a pod, Kubebay retries for a few seconds, then reports it as not evicted rather than deleting it. When it’s done, it says how many pods were evicted and skipped.

Before draining, check SPOF Radar for workloads a drain would take down.