Skip to content

Cluster consistency: hosts of one cluster that differ

The hosts of a cluster are meant to be interchangeable: any of them should be able to take any VM. Cluster consistency puts them side by side and warns about the one that drifted:

Migration to xcp-a2 will fail: its port lacks VLAN 20 (core-sw1 Gi1/0/6), which xcp-a1 carries.

It compares, host against host:

  • the hypervisor build (ESXi, Proxmox VE, XCP-ng);
  • the BIOS and BMC firmware, between hosts of the same model;
  • the MTU of each network the hosts share;
  • the VLANs the cluster's VMs use, against what each host's switch ports carry.

Differences are warnings, never critical: nothing is broken until a VM moves. Cluster consistency is part of Business and MSP. Collimo only reads: it never changes a host, a port or a setting.

Cluster consistency on the Virtualization tab: vCenter's Prod consistent; Proxmox VE's lab with its two nodes on different builds and BIOS, and no majority; the XCP-ng pool Prod with one host's BIOS and MTU, another's missing VLAN 20 and a third's older build marked

Before you start

  1. Your hypervisors: vCenter, Proxmox VE or XCP-ng, and an inventory. Hosts, their builds and their networks come from the newest inventory, which runs daily by itself on Business. To get one now, open the Virtualization tab and choose Take inventory now.
  2. Your servers' management controllers (Redfish), for the BIOS and BMC firmware (see Servers). A machine with no BMC but with Intel AMT, such as a mini PC with vPro, gives its BIOS through AMT (see Machines without a BMC). A host with neither shows not read.
  3. Your switches, for the VLANs: UniFi (with the console) and any others over SNMP (see Switches over SNMP).

Why the inventory, and not the live poll

XCP-ng only gives a host's build, and vCenter its switches' MTU, in the wider read the inventory makes. Reusing it costs nothing extra. A host patched since then shows the change after the next inventory.

Read the panel

Open the Virtualization tab and scroll to Cluster consistency.

  1. Each cluster (or XCP-ng pool) has its name on a label, and consistent or how many hosts differ.
  2. The table has one row per host: Build, BIOS, BMC, MTU and VLANs. A value marked ≠ differs from the rest of the cluster. It is amber until you mark its warning seen in Needs attention.
  3. The Compared line under the table says what each column found: same, differs, not compared (fewer than two hosts read) or ports not known.
  4. Under the table, each difference in a sentence, with what the other hosts have.

The same warnings are listed on the Resilience tab under Cluster drift.

Which one is wrong?

The value most hosts share counts as right, and the others are named. With no majority (two hosts, two builds), Collimo can't tell which side is right, so it names every host, each with the other side.

Fix a difference

Collimo doesn't change anything. For each warning:

  • Build: patch the odd host to the cluster's build, or the rest to its build, in a maintenance window.
  • BIOS or BMC: update the odd host's firmware from the vendor's tool to match the others.
  • MTU: set the odd host's switch, bridge or network to the MTU the others use. The physical switch port must allow it too.
  • VLAN: add the VLAN to the host's switch port (as a tagged VLAN, or in the port's profile on UniFi). The next poll sees it.

Ports not known

The VLAN check only runs when Collimo has found every host's switch ports, and read their VLANs. If it hasn't, the column says ports not known and nothing warns: a port Collimo can't see might well carry the VLAN. To fix it, see how Collimo finds where a host plugs in. A powered-off host is left out of the VLANs: it isn't a migration target.

What Collimo compares, exactly
  • Build: vCenter's product version and build; Proxmox VE's version and kernel; XCP-ng's version and build.
  • BIOS and BMC: from Redfish, or the BIOS alone from Intel AMT, matched to the host by serial number, then by name. A placeholder serial such as Default string isn't one, so those machines are matched by name. Hosts of different models aren't compared: their BIOS versions are different lines.
  • MTU: on vCenter, each standard and distributed switch, and each VMkernel port by its port group; on Proxmox VE, each bridge and bond by name (no mtu line counts as default); on XCP-ng, each pool network.
  • VLANs: on vCenter, the port groups of the cluster's VMs and VMkernel ports; on Proxmox VE, the VM network cards' VLAN tags; on XCP-ng, the VM networks' VLANs. Untagged and trunked traffic names no VLAN.
  • Not compared: network card firmware and drivers. vCenter, Proxmox VE and XCP-ng don't all report them in what Collimo reads.

Metric

infra_cluster_drift_hosts{source, cluster, check} is how many hosts differ on each check (build, bios, bmc, mtu, vlan) that compared two or more hosts.