Skip to main content
Nodes lists the machines in the cluster, with the ones in trouble first. Open it from the sidebar under Cluster, with G then N, or with ⌘5 in the desktop app (Ctrl5 on Windows and Linux). Nodes aren’t in a namespace, so the list always shows all of them.
The Nodes list: four nodes with their roles, CPU and memory meters and versions, one under memory pressure and one not ready.The Nodes list: four nodes with their roles, CPU and memory meters and versions, one under memory pressure and one not ready.

Nodes, their usage and conditions: one under memory pressure, one not ready.

The list

A node’s status is one of:

Open a node

Overview has:
  • Details: its internal IP, OS image, kernel, container runtime, kubelet version, architecture, pod CIDR, and whether it’s Schedulable.
  • Capacity: CPU and memory in use against what’s allocatable, and how many pods can be scheduled on it.
  • Taints, when it has any.
  • Conditions, labels and annotations.
Pods lists every pod running on the node, from every namespace, failing ones first. Metrics charts its CPU and memory over time against its capacity, when Lumovi has found your usage history. Map shows the pods that run on it and what runs them. See Map. Events and YAML are there as for any object. A node has no Logs tab: that would be everything on it.

A shell on the node

Shell, at the top of the panel, opens the node’s Shell tab. From there, Lumovi starts a short-lived privileged pod on the node and opens a shell through it: as root on the node itself, or in the pod, with the node’s files under /host. Nothing starts until you choose Start shell, and the pod is deleted when the shell ends. Your RBAC, and the Pod Security of the namespace the pod runs in, decide whether you can. See Shells on nodes.

Take a node out of service

A drain takes kubectl drain’s options:
  • Grace period: how long each pod has to stop before it’s killed, at least 1 second. By default, each pod’s own terminationGracePeriodSeconds (--grace-period).
  • Give up after: how long to keep trying pods a PodDisruptionBudget holds. Then those pods stay, and so does the cordon (--timeout).
  • Delete emptyDir data, for pods with emptyDir volumes, whose data is lost (--delete-emptydir-data).
  • Leave DaemonSet pods, on by default: their DaemonSet would start them here again anyway. Off, a DaemonSet pod stops the drain (--ignore-daemonsets).
  • Evict pods without a controller, for pods that nothing will recreate (--force).
Static pods stay where they are: the kubelet manages them. Drain stays disabled until the options the node’s pods need are ticked. Evictions respect PodDisruptionBudgets: a pod a budget holds is tried again whenever its budgets allow, its row saying which budget and why, until the drain gives up or you stop it. Then the dialog says which pods it stopped on, like “Stopped after 5:00: 2 pods are still on worker-2, held by PodDisruptionBudgets. It stays cordoned.” In a cluster named like production, you type the node’s name first.

Taints

Taints… lists the node’s taints, and adds and removes them. A taint keeps pods off the node unless they tolerate it, by its effect:
A node's taints without YAML: one to add and one to remove, staged, with the kubectl commands they are.A node's taints without YAML: one to add and one to remove, staged, with the kubectl commands they are.

A node’s taints: one to add and one to remove, with the kubectl commands they are.

Remove a taint with its Remove button, and Keep to take that back; add one with a key, an optional value and an effect. Nothing changes until Apply. Taints Kubernetes sets itself, like a node that isn’t ready, say Set by Kubernetes, and removing the control plane’s says that pods will then be scheduled on it. Where a NoExecute taint is staged, the dialog says first which pods it evicts at once, like “Applying this evicts 3 pods from worker-2 at once.”, and its button says so: Apply and evict 3 pods. Undo, in the notification, puts the taints back; pods it evicted don’t come back. Apply sends one change that touches only the taints you added or removed, each removal checked against the taint as Lumovi read it. If the node’s taints changed while the dialog was open, nothing is changed, and it shows them as they are now, with your changes still staged. The command beside it does the same with kubectl taint, and the audit log records it as resource.patch, like “Tainted worker-2 with dedicated=gpu:NoSchedule” or “Changed worker-2’s taints: dedicated=gpu:NoSchedule added, maintenance:NoExecute removed”.
The Drain worker-2? dialog, on production: the node is cordoned, then its pods are evicted, PodDisruptionBudgets respected. Delete local data and Evict pods without a controller are unticked, and 2 pods need one of them before the node can be drained: 8 pods to evict, 4 staying, kube-proxy among them as a DaemonSet pod.The Drain worker-2? dialog, on production: the node is cordoned, then its pods are evicted, PodDisruptionBudgets respected. Delete local data and Evict pods without a controller are unticked, and 2 pods need one of them before the node can be drained: 8 pods to evict, 4 staying, kube-proxy among them as a DaemonSet pod.

Draining a node under memory pressure: the pods it evicts, those that stay, and what has to be allowed first.

Pick several nodes with X to cordon or uncordon them together. More in Actions by kind.

Actions by kind

Every action, with what it runs and what it can undo.

Live usage

Where the CPU and memory figures come from.