Proxmox Backup Server now holds the lab’s virtual machine backups. It was built on 19 October as an interim server, with the long-term home still to come, and it is counted as in progress on the status page, not complete. This note is about why the backups moved, and about what “interim” means here.
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 old path was doing
Virtual machine backups used to stream the hypervisor’s native dump output into the file-level backup tool. That format does not deduplicate. Block order and headers shift between runs, so every weekly guest backup was effectively a full copy. The repositories grew into the terabytes on a NAS that had only a few terabytes free.
The new server reads only the blocks that QEMU’s dirty bitmap marks as changed, stores fixed-size chunks addressed by content, and keeps identical data once. On the proof run, the first backup of a small guest transferred its full size and the second transferred about sixty megabytes, with almost everything reused. Two snapshots occupied a small fraction of what two full copies would have.
Backups are also encrypted on the client before they leave the hypervisor, so a copy sitting on shared storage is ciphertext.
Two datastores, on purpose
The primary datastore is a local disk on the backup server’s own virtual machine. The second is a copy on the NAS, kept up to date by a daily sync job.
Garbage collection touches every chunk and verification reads every chunk. On a local disk that takes minutes, and over NFS to a busy NAS it takes hours and competes with every other host’s backup run. A local primary also keeps backups landing when the NAS is down: the server hard-requires the local mount and only softly wants the NFS one, so a dead NAS fails the sync job and nothing else.
The NAS copy is there so that the planned replication of the NAS to a second box carries the backups too, and so that the later move to the permanent server is a sync job, not a re-backup: add this server as a remote there, pull the datastore, repoint the storage definitions. Nothing is re-encrypted.
What “interim” costs
The server is a virtual machine on one of the hypervisors it backs up. A backup server that shares fate with what it protects is acceptable only as a stage. The permanent target is the storage host that is still being built, and the interim design exists so that the data can move there without starting over.
The first sync failed on a permission, not on a network
The NAS share’s root is readable only by root, and the backup service runs as an unprivileged user, so the first sync failed with permission denied on the chunk directory. The fix was to mount the subdirectory itself, so the client never has to look up through the restricted root. Nothing else on the share was changed.
What is proven and what is not
One manual restore has been proven. There is no scheduled restore drill for this server yet, and it is not on the restore timers the older backups use. The server’s coverage and staleness alerts exist. The status page lists it as in progress for exactly those reasons, and it will stay there until a drill runs on a timer and the first real restore has come from this server and not from a test guest.
What would change the decision
If the permanent storage host ends up unable to run the backup server, or if the interim server’s own failure ever costs a backup window, the interim design would need to be replaced, not extended. Until then, deduplication alone justified the move: the old path’s cost was growing faster than the storage it was writing to.