What a snapshot can and can’t do:
| Snapshot | vzdump backup | |
|---|---|---|
| where it lives | where the guest’s disk is | on another storage: NAS, a second disk, PBS |
| how long it takes | seconds | minutes to hours |
| rollback | seconds, qm rollback |
a restore, minutes |
| survives the disk dying | no | yes, if the storage is different |
| survives the guest being deleted | no: qm destroy deletes the snapshots too |
yes |
| what it costs | grows over time: keeps old versions of changed blocks | space — one copy per day of retention |
Hence the order for a risky change — an OS upgrade in the guest, a database version change, a network edit:
qm snapshot 101 before-upgrade --description "before upgrading to 26.04"
# make the change, check, live with it for a day or two
qm delsnapshot 101 before-upgrade # everything works — the snapshot isn't needed anymore
# it doesn't work:
qm rollback 101 before-upgrade
Why not keep snapshots for months:
- Space. On LVM-thin and ZFS a snapshot keeps the blocks that were overwritten after it. A VM with a database rewrites its disk several times a month, and the snapshot grows to the size of the disk. On a thin pool that’s the road to 100 % full (Proxmox storage).
- Speed. With qcow2 every snapshot is one more link in the chain, and reads go through all the links.
- On ZFS you can roll back only to the latest one. To go back to a month-old snapshot you have to delete every newer snapshot.
- A rollback returns everything. Along with the bad change, you lose the data the VM wrote after the snapshot: mail, database records, uploaded files.
Before a change that touches data, do both: a snapshot for a quick rollback and a fresh backup — Backup now on the VM — in case you have to roll back from another host (Backups in Proxmox).
How to check
qm listsnapshot 101 # only fresh snapshots, each with a clear reason in the description
lvs pve | grep snap # snapshot volumes on LVM-thin and how much they take