The first guest built on a chassis blade came up fine and answered on the network. The blade it ran on disappeared. No ping, no SSH, no management web interface, while the chassis still reported the blade as powered on. A running, reachable virtual machine on a host nobody could reach reads like a firewall or routing fault. It was neither, and the agent had caused it on purpose-built test hardware.
Drafted by an AI agent (claude-opus-5) from this lab’s own runbooks, deployment log and errata. Reviewed before publication by the site owner. How this site is written →
What the host’s network looked like
The blade’s bridge carried the host’s own address, and a second interface stacked on it carried the host’s address on the lab network, tagged for the lab VLAN. The bridge was not VLAN-aware, which in Proxmox terms means VLAN filtering was switched off.
The new guest asked for a network interface on that bridge, tagged for the same lab VLAN. On a bridge that is not VLAN-aware, Proxmox does not refuse the tag. It falls back to an older mechanism: it creates a new sub-bridge for that VLAN and moves the physical port into it.
That left the host’s tagged address stacked on the original bridge, which no longer had the physical port. The address was attached to nothing. The guest worked precisely because it sat on the new sub-bridge that had taken the port.
Recovery, and the control that made it readable
A power cycle fixed it, since the sub-bridge exists only while the guest is running and is never written to the network configuration file; a reboot restores the host as it was. It took about six minutes on this hardware. A second blade stayed up throughout and served as a control, because without one, “this blade is down” could not have distinguished a blade fault from a chassis or switch fault.
The check that did not exist
Before creating a guest, verify that the target’s bridge is VLAN-aware by reading it, not by inferring it from the bridge existing, and not from another guest already being tagged on a different host.
cat /sys/class/net/vmbr0/bridge/vlan_filtering # 1 = safe to tag, 0 = do not tag
A zero means either make the bridge VLAN-aware or do not tag at all. This is now a required preflight for the guest builder, in the same class as the guard against reusing a machine ID: a check that runs before the destructive step, not a note in a design document.
The fix to the build itself
The blade’s first-boot configuration now sets the bridge VLAN-aware, with the full VLAN range, so a rebuilt blade can host tagged guests. It was proven on hardware by rebuilding a blade and repeating the exact action that had caused the outage. A guest with the tag started, the sub-bridge was never created, the physical port stayed on the original bridge, and the host kept its lab address. The blade reported that over its own network connection, which is the only evidence that means anything here, since a host that has lost its address cannot say so.
The preflight is still the real guard
A blade installed before the fix, or rebuilt by hand, still has the unsafe bridge, and nothing else would notice. The preflight is what protects those, which is why the fix to the build did not replace it.
The wider class
This was the second VLAN-layer trap in two days on the same feature, and both had the same shape. A property that looks like it belongs to the bridge actually belongs to a layer underneath it. The day before, a VLAN-aware bridge was configured correctly while the uplink switch port did not carry the VLAN, so frames died one layer down. This time the bridge was not VLAN-aware, and asking for a tag silently restructured the layer beneath it. VLAN filtering being on and the switch trunking the VLAN are two independent facts, neither of which shows in the guest’s configuration, and each has to be read on the specific host.