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.
A backup store is bigger than the server it protects when the history kept, the data that does not shrink, and the space not yet reclaimed add up to more than one copy of the source. It is not a fault. It is arithmetic, and each piece can be measured separately. If a provider reports a figure larger than your data, the useful question is which of five pieces is responsible.
Below, the footprint of one server is added up step by step, followed by how to read the figures your own tools give you.
Why would a backup be larger than the thing it backs up?
Because a backup store holds more than today's copy. It holds the first full copy of the data, every change kept since, and anything that has been removed from the retention list but not yet physically freed. Deduplication keeps repeated data from being stored twice. It does not remove the first copy, and it does not remove changes that are genuinely new.
Microsoft's Data Deduplication overview describes backup snapshots as having "minor differences from day to day", and says that datasets with high duplication could see "optimization rates of up to 95%, or a 20x reduction in storage utilization". That is a ceiling for very repetitive data, not a promise. It also describes savings on a volume, not the cost of keeping a long history.
The five pieces of one server's footprint
| Piece | What it is | What makes it large | Where you can see it |
|---|---|---|---|
| 1. First full copy | Everything is stored once on the first run | Data that does not compress: photos, video, archives, already-encrypted files | Size of the first run |
| 2. Changed data kept | New or changed blocks from each later run | Churn: databases, mailboxes, logs, virtual disks | Size added per run |
| 3. Retention | How many days of changes are kept | Longer retention keeps every day's changed data longer | Policy settings |
| 4. Unreclaimed space | Data no longer referenced but not yet freed | Cleanup that is scheduled, skipped or limited | Repository size against the sum of what is referenced |
| 5. Whole-chunk rounding | A small edit re-stores a whole chunk | Large chunks, big files with scattered edits | Size added per run against bytes you changed |
The sections below take each piece in turn.
Does compression shrink the first copy?
Only the part that can shrink. Backup tools compress what they store, but photos, video, zip files and encrypted files are usually already compressed or look random, so the tool gets little or nothing back.
Borg documents this directly. Its default compression is lz4, described as "Very high speed, very low compression". It also has an auto mode whose heuristic "tries with lz4 whether the data is compressible" and "for incompressible data, it will not use compression". The page says this "can be helpful for media files which often cannot be compressed much more".
restic works differently by version. Its documentation says repository format version 2 "adds support for compression", and that after upgrading an older repository "File contents stored in the repository will not be rewritten, data from new backups will be compressed". A repository created before format version 2 can therefore hold uncompressed data from earlier backups.
If a provider quotes a saving from compression, ask what share of your data is compressible. A file server full of scanned PDFs and project videos will not behave like a server of text documents.
Does each day's backup add a full copy?
No. With a deduplicating store, a run stores only chunks it has not seen. restic's backup documentation shows a repeat run where "no new data was added to the repository (since all data is already there). This is deduplication at work!" Borg says the same by another route: with --stats, "the 'This Archive' deduplicated size there is most interesting as that is how much your repository will grow".
So the growth per run is the amount of new or changed data, not the size of the server. That is also why a nightly "full" backup in a deduplicating store does not cost a full copy each night. Its cost depends on how much data changed.
Why does a small edit cost more than the bytes you changed?
Because stores keep data in chunks, and a changed chunk is stored again as a whole. restic's design reference says "Files smaller than 512 KiB are not split, Blobs are of 512 KiB to 8 MiB in size" and that "the implementation aims for 1 MiB Blob size on average". Borg's default chunker parameters give a 2 MiB average, between 512 KiB and 8 MiB.
If a large database file has a few bytes changed in many scattered places, each edit can land in a different chunk, and every chunk that was touched is stored again in full. That is the "whole-chunk rounding" in the table. Compression then shrinks the stored chunks, so the real figure is smaller. The bigger the chunk, the more a small edit costs, but the fewer chunks the store has to index.
What does deleting old backups free, and when?
Not always straight away. In restic, removing a snapshot with forget does not free space on its own. The documentation says "the data that was referenced by files in this snapshot is still stored in the repository", and "to cleanup unreferenced data, the prune command must be run". Even then, prune accepts --max-unused, which allows "unused data up to the specified limit within the repository", with a default value of 5%. A restic repository can therefore legitimately carry up to about 5% of unused data after a prune.
Borg has the same two-step shape. The compact command "frees repository space by compacting segments", and the page says "only compaction will really free repository space". A segment is compacted only if the saved space is above the threshold, which is 10% by default.
Risk: forget, prune, delete and compact permanently remove stored data. There is no undo once the unreferenced data is gone. restic also locks the repository during a prune: "During a prune operation, the repository is locked and backups cannot be completed." Run clean-up only when you are sure which snapshots you intend to lose and no backup is due.
What does one server add up to?
Example (illustrative): a file server fs01 holds 800 GB. Of that, 600 GB is documents and databases that compress to 75% of their size, and 200 GB is photos and archives that do not compress. The backups go to /srv/backup on backup01, keeping 30 daily restore points. The figures are made up to show the arithmetic.
- First full copy: 600 GB × 0.75 = 450 GB, plus 200 GB uncompressed, which is 650 GB.
- Daily change: 2% of 800 GB is 16 GB new each day. The same mix applies: 12 GB compressible × 0.75 = 9 GB, plus 4 GB that does not compress, which is 13 GB stored per day.
- Days of change kept: the first full copy is the oldest of the 30 restore points, so 29 days of change sit on top of it: 29 × 13 GB = 377 GB. This assumes each day changes different blocks, so nothing is changed twice.
- Unreclaimed space: up to 5% of the store, which is 5% × (650 + 377) GB = 51 GB, rounded.
- Total: 650 + 377 + 51 = 1,078 GB, against 800 GB of source data.
The store is 1,078 ÷ 800 = about 1.35 times the data it protects. Nothing has gone wrong. The 278 GB of extra is the 377 GB of retained history plus 51 GB of unreclaimed space, less the 150 GB that compression saved on the first copy.
Now change one input. If retention doubles to 60 daily points, the history becomes 59 × 13 = 767 GB and the total grows to 650 + 767 = 1,417 GB before unreclaimed space. Retention, not data size, moved the bill.
How do I check this on my own system?
Work from measured figures, not from a provider's headline ratio.
With restic (read-only commands; both are documented as counting modes of stats):
restic stats --mode raw-data
restic stats --mode restore-size latest
The manual describes raw-data as "Counts the size of blobs in the repository, regardless of how many files reference them", and restore-size (the default) as "Counts the size of the restored files". The manual does not say whether raw-data is counted before or after compression, so compare it with the storage's own usage figure for the repository folder.
With Borg:
borg info
The documentation says "the deduplicated sizes of the individual archives do not add up to the deduplicated size of the repository ('all archives'), because the two mean different things". "This archive" is data stored only for that archive. "All archives" is the amount stored in the repository. Use the "All archives" deduplicated figure for the bill.
On a Windows volume that holds the store, the volume figures from Get-Volume give Size and SizeRemaining. Used space is Size minus SizeRemaining.
Questions to ask a provider:
- Is the figure on the invoice the size of the source data, the size of all restore points added up, or the space the store actually takes?
- How long is each class of data kept, and is the kept history counted?
- Is space freed by the clean-up job counted on the invoice only after it runs, and how often does it run?
- What compression and chunk size are used, and what share of my data does not compress?
- Is a second copy elsewhere counted in the same figure?
Read next
Related articles you may want next.
- Backup & RecoverySize Backup Retention Using Measured Storage GrowthBuild a capacity forecast from stored growth, copy formats and retention dependencies, with explicit headroom, recovery needs and uncertainty.Read article
- Backup & RecoveryGrandfather-father-son retention: counting the restore points a policy really keepsHow to count the restore points a 30 daily, 12 monthly backup policy keeps, where the gaps fall, and what to ask a provider to confirm in writing.Read article
- HostingThin or thick provisioned storage: what you pay for and what can run outWhat thin and thick provisioning mean on a hosted VMware server, why a full shared datastore can stop a server whatever its disk format, and the questions to ask about billing and monitoring.Read article
Risks and limits
- These commands have not been run or lab-tested. They are taken from restic, Borg and Microsoft documentation.
- The worked example uses round, made-up figures and assumes unique daily changes. Real churn overlaps, and real compression varies by file type.
- The restic and Borg pages show no dates. Borg's stable pages are labelled version 1.4.5. Borg 2.x is a different generation of the tool and its commands differ.
- The 5% and 10% figures are tool defaults, not guarantees. A provider may use different tools and settings.
- Microsoft's 95% and 20x figures describe the best case for very repetitive data, not a typical backup store.
- How a provider shares stored data between customers, and how cloud object storage charges per request, change the bill separately from the sums above.
NetLeverage includes a daily full disaster-recovery backup of each hosted server at no extra charge. Talk to an engineer.
Sources
- Data Deduplication overview: Microsoft Learn, page date 2025-06-20, accessed 17 November 2026
- restic design reference: restic documentation 0.19.1, no page date, accessed 17 November 2026
- restic: removing backup snapshots (forget and prune): restic documentation, no page date, accessed 17 November 2026
- restic: backing up: restic documentation, no page date, accessed 17 November 2026
- restic: working with repositories: restic documentation, no page date, accessed 17 November 2026
- restic-stats(1), v0.19.1: restic project, accessed 17 November 2026
- borg compact: Borg 1.4.5 documentation, no page date, accessed 17 November 2026
- borg create: 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 help compression: Borg 1.4.5 documentation, no page date, accessed 17 November 2026
- Borg internals: data structures: Borg 1.4.5 documentation, no page date, accessed 17 November 2026
- Get-Volume: Microsoft Learn, Windows Server 2025 PowerShell reference, accessed 17 November 2026
This article was originally published on .