About this article: This article is reviewed against the cited sources. This is source research, not lab validation. Check the stated versions and your own requirements before making changes.
Both answers are correct, about different things. What a deduplicated backup restores is a full copy: each backup describes the whole data set as it stood that day. What it stores is incremental: only pieces of data the store has not seen before are written. A restore reads the backup's list and every stored piece that list points to, and some of those pieces were written by backups from weeks or months earlier.
That one fact explains why backups are small, why deleting an old one frees less than expected, and why a single damaged piece can matter to many backups.
What does "full backup, incremental storage" mean?
Three open-source backup tools describe this design in their own words.
Proxmox Backup Server documents that when a client uploads a snapshot, it "can send only the checksum instead of data and checksum" for a chunk the server already has. It then states: "This way the actual upload of Snapshots is incremental while each Snapshot references all chunks and is thus a full backup."
BorgBackup's FAQ says the same in fewer words: "Borg always creates a new full archive, but only transfers and stores new/changed chunks."
restic splits files into variable-length pieces it calls blobs. Files smaller than 512 KiB are not split; blobs run from 512 KiB to 8 MiB, with an average aim of 1 MiB. For a modified file, "only modified Blobs have to be saved in a subsequent backup", even when bytes are inserted or removed in the middle.
In all three, the unit of storage is a piece of data identified by a hash of its content. The unit a person sees (a snapshot or archive) is a list of those pieces.
What does a restore actually need?
A restore needs the list for the backup you chose, plus every piece on that list. Nothing in the design requires the pieces to come from that day's backup. A piece first written in January is used by every later backup in which that data did not change.
restic documents the dependency rules that keep a store usable. In its repository format, "a snapshot must only reference an existing tree blob", a reachable tree must only reference blobs that exist, and an index must only reference blobs in existing packs. restic says these are strict requirements, and that breaking them "will lead to data loss".
This is different from a classic chain, where a restore needs the last full backup plus each incremental since. In a deduplicated store there is no chain to follow, but there is a web of shared pieces. For the person doing a restore, the difference shows up in the questions they must ask:
| Question | Classic full plus incrementals | Deduplicated store |
|---|---|---|
| What does "incremental" describe? | What each backup file contains | What was uploaded or written this time |
| What does a restore read? | The full backup and each later increment up to the point | The chosen backup's list and every piece it references, whichever backup wrote them |
| What can one damaged item affect? | Later increments in that chain | Every backup that references that piece |
| How is an old backup removed? | Delete files, subject to later increments depending on them | Remove the list, then reclaim pieces nothing else references |
| What is the size of one backup? | Its own file | Ambiguous: its full size or what only it stores |
How does deleting an old backup work?
Deleting a backup in a deduplicated store has two stages, and storage is freed only in the second.
Proxmox Backup Server says that when pruning a snapshot, "only the snapshot metadata (manifest, indices, blobs, log and notes) is removed. The chunks containing the actual backup data and previously referenced by the pruned snapshot, have to be removed by a garbage collection run." Even then, the no longer referenced chunks "will then be marked for deletion on the next garbage collection run and removed by a subsequent run after the grace period".
restic splits it the same way: forget removes snapshots, and prune removes "the remaining data that was referenced only by the removed snapshots". The restic design page states: "Only the prune operation removes data from the repository."
Borg's delete page warns: "When deleting archives, repository disk space is not freed until you run borg compact."
The consequence is practical. Deleting the oldest backup frees only the pieces no remaining backup uses. Pieces that later backups reuse stay. Proxmox adds a privacy note: sensitive data in a chunk "will outlive pruned snapshots and remain present in the datastore as long as referenced by at least one backup snapshot".
Example (illustrative): fs01 holds 1,000 GiB. Assume 1 percent of its data is new or changed each day, 10 GiB, and no compression. The first backup stores 1,000 GiB. Each of the next 29 nightly backups stores about 10 GiB of new pieces. After 30 backups the store holds about 1,000 + 29 × 10 = 1,290 GiB. Thirty separate full copies would take 30 × 1,000 = 30,000 GiB, about 23 times more. Now delete the oldest backup. It frees only the pieces that no later backup still references: those that later changes have replaced. If the 29 days of changes each replaced different data, that is at most 29 × 10 = 290 GiB, so the store shrinks from 1,290 GiB to about 1,000 GiB, not by the 1,000 GiB the backup appears to contain. If the same data changed every day, far less is freed. And on any day, a restore of backup number 30 reads about 1,000 GiB of pieces, most written by the first backup.
What does a damaged piece do?
Because backups share pieces, one damaged stored piece can affect every backup whose list points to it. This follows from how the design shares data. Proxmox documents that "multiple indexes can reference the same chunks", and restic says "All files in a repository are only written once and never modified afterwards", which is why a store can detect accidental changes by comparing each file's hash with its name.
A backup can therefore look intact, with a complete list and a successful job report, while a piece it needs is unreadable. The only way to know is to read the pieces back. What each tool's check proves, and what only a restore proves, has its own topic: "What restic and Borg repository checks prove, and what only a restore can".
How can you check this on your own system?
Ask the provider or read your own tool's output for the following.
- How much is stored only for each backup. Borg documents that in
borg info, "This archive / deduplicated size = amount of data stored ONLY for this archive = unique chunks of this archive", while the "All archives" figure is the whole repository. It warns that the per-archive figures do not add up to the total.borg info /path/to/repo --last 1is a read-only command in Borg's own examples. - What will deleting free. Borg's
deletehas--dry-run --listto "see what would be deleted" and--stats, whose "Deleted data" deduplicated size "is how much your repository will shrink". The two options are mutually exclusive. - How your provider counts "size". Ask whether the bill uses the full size of each backup, the unique data, or the total stored. Without that answer you cannot predict what deleting backups will save.
- When clean-up runs. On Proxmox, pruned data waits for garbage collection and a grace period. Ask how often that runs and who can run it.
Disruptive step: borg delete with borg compact, restic forget --prune and Proxmox prune with garbage collection permanently remove backups and, later, their unshared pieces. After the space is reclaimed there is no way back from those commands. Run a dry run first where the tool offers one, and test a restore of a backup you intend to keep.
Read next
Related articles you may want next.
- Backup & RecoveryWhat restic and Borg repository checks prove, and what only a restore canWhat restic check, --read-data, borg check and --verify-data actually prove, what they cost, and the restore test that covers what they cannot.Read article
- Backup & RecoveryBacking up and restoring Data Deduplication volumes: what is really on the diskPlan Windows Data Deduplication recovery around the chunk store, supported backup modes, compatible servers and realistic restore capacity.Read article
- Backup & RecoveryFinding a bad backup before the day you need itWhat hashes, signatures and sample restores detect in backups, which faults they miss, and when to treat a copy as unusable.Read article
Risks and limits
- These commands have not been run or lab-tested. They are taken from each project's documentation.
- Proxmox Backup Server, restic and Borg are examples of the design. Other products may chunk, index and clean up differently, and a vendor's own documentation is the source for its product.
- The documentation pages show no dates. They name themselves Proxmox Backup 4.2.11-1, restic 0.19.1 and Borg 1.4.5, and other releases may differ.
- The statement that a damaged piece affects every backup that references it is not quoted from a vendor page; it follows from the documented sharing.
- Deduplication reduces storage, and it does not replace a second copy. A damaged store holds every backup in it.
- The example's 1 percent daily change and the absence of compression are assumptions for arithmetic, not typical figures.
Sources
- Technical Overview (datastores, snapshots, chunks) — Proxmox Backup 4.2.11-1 documentation, no page date, accessed 17 November 2026
- Maintenance Tasks (pruning, garbage collection) — Proxmox Backup 4.2.11-1 documentation, no page date, accessed 17 November 2026
- Design: repository format, ordering rules, backups and deduplication — restic 0.19.1 documentation, no page date, accessed 17 November 2026
- Removing backup snapshots — restic 0.19.1 documentation, no page date, accessed 17 November 2026
- Frequently asked questions — Borg 1.4.5 documentation, no page date, accessed 17 November 2026
- borg info — Borg 1.4.5 documentation, no page date, accessed 17 November 2026
- borg delete — Borg 1.4.5 documentation, no page date, accessed 17 November 2026
This article was originally published on .