An incremental backup holds only the disk blocks written since the backup before it. A chain is one full backup followed by up to N incrementals; then the next backup is full again and starts a new chain. Incrementals work on SFTP, S3, NFS and local destinations, with encryption, Object Lock and Verify. They are off by default: every backup is full until you set a number.
Turn them on#
Set Incremental backups per full backup (0 to 30) wherever you set a schedule: the default schedule under Backups › Schedules, a plan's own schedule, or a server's own schedule. 0 means every backup is full. With daily backups, 6 gives one full backup a week and six incrementals.
Existing schedules keep 0 until you change them.
Which servers get incrementals#
The node makes an incremental only when it can prove its record of the previous backup is complete. A server gets a full backup instead when:
| Situation | Why |
|---|---|
| The server is stopped | The change tracking (libvirt checkpoints with dirty bitmaps) needs a running or paused server |
| A disk is raw, ZFS, LVM-thin or Ceph RBD | Change tracking is kept only in qcow2 disks |
| The destination is a Proxmox Backup Server | PBS deduplicates by itself; every PBS backup is a full image that uploads only new chunks |
| A disk was added, deleted, resized or moved | The disk set no longer matches the previous backup |
| A snapshot was taken or reverted, the server was reinstalled, restored, rescued or moved to another node | The disks may have changed outside the running server |
| QEMU or the host crashed | The change tracking is marked inconsistent |
| The previous backup failed, was deleted or is on another destination, or the chain is full | The panel asks for a full |
When the panel asked for an incremental and the node made a full instead, the backup records why (chain_reset), and the audit log has a backup.chain_reset entry with the reason.
Containers take incrementals too, by comparing 1 MiB blocks of their volume with the previous backup; a container whose volume was resized gets a full.
How chains are stored#
Each backup of a chain is its own directory on the destination, laid out like any other backup. An incremental's disk file is a qcow2 holding only the changed clusters, pointing at the previous backup's file. Incrementals are not compressed (they hold only what changed); full backups are.
On the server's Backups tab each backup shows its type ("Full", "Incremental 2 of 6") and the size With its chain. Backups › Restore has a Type column.
Keeping and deleting chains#
Chains are kept and deleted whole:
- Retention still counts restore points ("Backups kept"), but deletes a chain only when none of its backups is among those kept. With Backups kept at 7 and 6 incrementals per full, up to 13 backups may be stored for a while.
- Delete older than (days) applies to a chain's newest backup.
- Deleting a full backup deletes its whole chain; the dialog says how many incrementals go with it.
- An incremental can be deleted by hand only when it is the newest of its chain. Otherwise the panel answers "Delete the newer backups of this chain first" (the API answers 409
chain_member). - A backup that a later backup builds on is never deleted while that later one exists. A kept backup, a locked one or one being restored holds its whole chain.
Restoring an incremental#
Restore and Restore to a new server work the same as for a full backup. The node downloads the full backup and each incremental up to the chosen one, checks each file's SHA-256, and merges them into one disk before swapping it in. The Restore dialog says what it downloads ("Restores the full backup of 1 Dec and 2 incrementals"). The node needs free space for the restored disk plus the largest incremental. A restore is refused if a backup of its chain is missing or being deleted.
Verify on Backups › Restore checks every file of the chain and that each incremental points at the right parent.
Requirements#
Nodes need an agent that reports incremental backup support; the panel sends incremental requests only to such nodes (older ones make full backups). Restoring or verifying an incremental also needs such a node. See Updating the agent.