Product Tour: Restarting a VM After Its Host Fails

A host stops answering. The question is not only “is the VM running again?” but “can people reach the service again, and is it safe — without two copies writing to the same disk?” This tour follows one failover from detection to a verified login.

Version: pre-release builds of 2026-10-14/15, shipped in release 81fce3de6Plan: BusinessTest dates: 2026-10-14 (Proxmox VE), 2026-10-15 (KVM)Environment: lab: two or three KVM hosts and two nested Proxmox VE hosts; VM disks on a distributed storage test clusterMaterial: recording and screenshots pending

Walkthrough

Detect. The host’s heartbeat stops; Nodekeeper collects evidence from more than one source before it decides the host is down, not just cut off from the master.
Decide. In confirm mode an administrator confirms in the panel; in automatic mode the controller proceeds on its own when its safety checks pass.
Fence. The old host is blocked from the VM’s disk on distributed storage, so the old copy cannot keep writing.
Restart. A healthy host is chosen and the VM is cold-started there with the same IP and MAC address.
Verify. The VM is checked until it answers; in the KVM run, SSH to the same address worked again.
Reconcile. When the failed host comes back, its stale VM definition is quarantined before fencing is lifted.

Screenshot pending — D6-a Failover timeline in the panel: suspected, evidence, fenced, restarted, verified A real, irreversibly redacted capture from a test environment will be placed here

Recording pending — D6-b Host frozen → VM back on another host → login to the same address A real, irreversibly redacted capture from a test environment will be placed here

Test record

VM HA restarts supported VMs (KVM, PVE) on another host with the same IP and MAC; measured 65 s (PVE) to about 4–8.5 min (KVM) in tests.

Test record
KVM, three hosts (2026-10-15)	host frozen → evidence complete after 93 s → administrator confirmed → fencing effective 12 s later → VM restarted on another host and verified: 3 min 50 s in total, including about 40 s of manual confirmation
KVM, two hosts (2026-10-15)	earlier run 8 min 26 s; after a detection fix, 4 min 21 s including 74 s of manual confirmation; SSH to the VM worked afterwards
Proxmox VE (2026-10-14)	frozen host → one-click confirmation → VM running on the second host in 65 s; the old host reconciled and fencing was lifted automatically
Proxmox VE, automatic mode (2026-10-14)	5 min 56 s from the last heartbeat; a second fencing of the same host within 24 hours put the group on hold until an administrator resumed it
Safety cases	two hosts frozen at the same moment → failover paused instead of acting; only the master-to-host link cut → treated as a network partition, no failover
Manual intervention	confirm-mode runs include the administrator’s confirmation time, as stated above
Test VMs	small Linux VMs; times on your hardware, network and VM count will differ

Different failures, different results

Recovery by failed component
What fails	What happens today	What is not promised
A KVM or Proxmox VE host running HA VMs	Fence, then cold restart on a healthy host with the same IP and MAC (steps above)	Live migration; anything only in memory is lost; data already written to disk is kept
The site node serving iSCSI to Windows office PCs	No automatic takeover yet. Windows PCs lose their disk until the node is back or replaced; running Linux PCs keep their disk because they map distributed storage directly.	Uninterrupted Windows sessions
An office PC’s own hardware	The user’s cloud PC can be started on another compatible PC after the old one is fenced	Moving open programs or unsaved work from the failed PC

Limits and prerequisites

Supported: KVM and Proxmox VE VMs whose disks are on Nodekeeper-managed distributed storage, in an HA group with spare capacity.
Not covered: ESXi, LXC containers, Proxmox corosync clusters, hyperconverged hosts, VMs on a LAN pool or behind NAT.
If the service is published through a CDN site or port forwarding on another node, verify that path separately — it is not part of this test record.
HA is not sold through WHMCS.

Read more: High availability · Infrastructure

← All tour topics Next: Secure hosting and per-site access Request sandbox access

The sandbox is invite-only and not open yet; there is no public live demo. Please do not send passwords, keys or customer data with a request.

N Nodekeeper

All-in-one Linux server management agent.

Product

Features Pricing Docs

Company

About Contact 中文

Support

[email protected] WhatsApp: +817093066006