Which physical disk is your virtual machine actually on?

Ask where a virtual machine's disk lives and the interface answers quickly: local-lvm. That is a true answer and it is not the answer you needed. local-lvm is a name, and a name does not tell you which piece of hardware stops turning.

The question worth asking is the one behind it: if one physical disk fails tonight, how many machines go with it?

The chain has five links

A virtual machine does not sit on a disk. It sits on the end of a chain, and every link is a place where the answer can surprise you.

The machine declares a virtual disk in its own configuration. A line like scsi0: local-lvm:vm-100-disk-0,size=32G names the storage, not the hardware.

The virtual disk is a file or a block device. Which of the two depends on the storage type, and that difference decides whether a snapshot is cheap or impossible.

The storage is an entry in the host's storage configuration. It is a label pointing at something else. Two labels can point at the same underlying thing, which is exactly how two machines that look separated end up sharing a failure.

The pool or volume group is where several physical disks are combined, or where a single one pretends to be a pool.

The physical disk is the only link that can actually die.

The list view shows you link one and link three. Links two, four and five are where the risk lives.

Why the answer matters more than it looks

Consider a common setup: eight machines, all on local-lvm, on a host with four disks. It reads like the load is spread across four disks. Whether it is depends entirely on how the volume group was built.

If those four disks were combined with striping and no redundancy, one disk failing takes all eight machines. If they were mirrored in pairs, one disk failing takes nothing and you replace it calmly. Same eight machines, same storage name, same screen. Opposite consequences.

Nothing in the machine list distinguishes these two cases. You have to walk the chain.

Where the chain hides itself

Thin provisioning. A machine can show 32 GB assigned while occupying 4 GB. Add up the assigned sizes and you can exceed the physical capacity without any warning appearing. Everything works until the pool actually fills, and then it fails for every machine at once rather than for the one that grew.

Snapshots on the same pool. A snapshot protects you from a mistake inside the machine. It does not protect you from the disk under it. A snapshot stored on the same pool as the machine dies with the machine.

Two labels, one device. A directory storage pointing at a path that happens to sit on the same volume group as a block storage. Two entries in the interface, one physical fate.

The boot disk. On many hosts the system disk and the first machine storage are the same device. It works, it is common, and it means a single failure takes both the machines and the thing that was supposed to manage them.

The question a backup does not answer

"There are backups" is a good answer to a different question. A backup answers how long recovery takes, not how many machines need recovering.

Those are separate numbers and they cost separately. Restoring one machine from a backup is an evening. Restoring eight is a weekend, and it is a weekend where nothing runs.

Walking the chain gives you the second number. It is worth knowing before you need it, because you cannot measure it while the disk is already dead.

Reading the chain by hand

Every link is readable from the host, and it is worth doing at least once so the shape becomes familiar.

The machine's configuration names its storage. The host's storage configuration says what kind of storage that is and what it points at. For a volume group, the group listing shows which physical devices belong to it. For a ZFS pool, the pool status shows the devices and, importantly, how they are arranged: whether they mirror each other or merely sit side by side.

That last detail is the one that decides everything, and it is the one furthest from the machine list.

What Atlas does

Atlas draws the whole chain on one screen, from the machine down to the physical disk, so the question is answered by following a line rather than by opening five screens and joining them in your head.

The map shows the machine, its processor and memory share, its virtual disks, the storage those disks sit on, the pool, and at the bottom the real hardware. When eight machines come down to the same physical disk, the eight lines converge visibly. You do not have to suspect it in order to find it.

The arrangement is drawn too, not just the membership: a mirrored pair and two independent disks look different, because the difference between them is the difference between a calm replacement and a lost weekend.

The chain is read live rather than drawn once, so a storage added last month is on the map without anyone having to remember to update a diagram.