The Data Protection Review
// Section
Linux-Backups
// Published
November 17, 2026
// Read time
8 min
// Share
LinkedIn
Linux-Backups

Kopia for Linux Backup: Content-Addressable Repositories and How It Compares to Restic and Borg

Kopia builds on content-addressable storage with encrypted packs, policies and a repository server. How it differs from restic and Borg, and when each fits.

The editors of The Data Protection Review
Editorial Staff
Close view of a server rack with glowing fiber cables and a laptop terminal in soft focus, suggesting encrypted snapshot storage
Close view of a server rack with glowing fiber cables and a laptop terminal in soft focus, suggesting encrypted snapshot storage

The usual shortlist for command-line Linux backup has two names on it: restic and BorgBackup. Our practitioner’s guide to Linux backup tools covers both. A third tool, Kopia, belongs in the comparison. It uses the same core idea, a content-addressable repository that stores each unique chunk once, but makes different choices about policies, interfaces and how clients reach the storage.

This piece explains how a Kopia repository works, and where its design differs from restic and Borg. It is a comparison of architectures, not a ranking. Details below come from the projects’ own documentation; the commands are copied from the Kopia docs and marked untested because we did not run them for this article.

What “content-addressable” means in practice

Kopia’s architecture documentation describes a block layer that gives every stored block an identifier derived from a cryptographic hash of its contents. Two identical blocks produce the same identifier, so a block that already exists is never uploaded twice. That is deduplication as a side effect of the addressing scheme rather than a separate feature.

Several details follow from that design:

  • Files are split into chunks first. Kopia’s compression documentation says files are cut with a configured splitter, each chunk is hashed to detect duplicates, and only new chunks are compressed and encrypted. The splitter used in the project’s own comparison is DYNAMIC-4M-BUZHASH, which uses content-defined boundaries so that an insertion in the middle of a file shifts only nearby chunks.
  • Small blocks are bundled into packs. The architecture page describes packs of roughly 20 to 40 MB, written with random names, so object-storage requests stay few and large. Prefixes distinguish data packs, metadata packs and index files.
  • An index maps blocks to packs. Each pack also carries its own small index at the end, so the repository can be recovered if index files are damaged.
  • Encryption comes after hashing. The default cipher is AES-256-GCM, with a master key derived from your passphrase using scrypt. Lose the passphrase and the data is gone, which the project says plainly.

Restic follows the same family of ideas. Its design documentation describes Rabin-fingerprint chunking with blobs between 512 KiB and 8 MiB, averaging about 1 MiB, every repository file named by its SHA-256 hash, and encrypted packs. Borg also chunks and deduplicates, and its FAQ describes a local cache of chunk checksums that drives the duplicate check. Underneath, the three are close cousins.

Where Kopia differs

Policies are a first-class object

In restic and Borg, retention and exclusion rules typically live in the command line or a wrapper script. Kopia stores policies, with inheritance down the directory tree, alongside the snapshots. The getting-started guide shows the pattern:

kopia repository create filesystem --path /tmp/my-repository
kopia snapshot create $HOME/Projects
kopia policy set --keep-weekly 30 .

(Commands from the Kopia documentation; untested here.) Compression is also a per-policy setting. Kopia’s documentation lists s2, zstd, pgzip, deflate and gzip variants, and recommends excluding file types that are already compressed, such as video, from compression to save CPU. It also notes that LZ4 is deprecated and that newer versions cannot read LZ4-compressed data, so an older repository needs migration before an upgrade.

More storage backends, and a repository server

Kopia can write to S3 and S3-compatible storage, Azure Blob Storage, Backblaze B2, Google Cloud Storage, WebDAV, SFTP and local or network filesystems, with Rclone-based and Google Drive support marked experimental. The S3 command on the repositories page is:

kopia repository create s3 --bucket=... --access-key=... --secret-access-key=...

(Also untested here.) The page says object locking can only be enabled from the command line when creating the repository, not from the GUI. That matters if you intend to pair Kopia with the object lock approach covered in our S3 immutability guide.

Kopia also offers a Repository Server, a way for clients to reach the repository through a server you run rather than holding storage credentials themselves. We did not find the server’s access-control details in the pages we read, so check its documentation before you rely on it for tenant isolation.

A GUI exists

KopiaUI is a desktop front end for the same repository format. For an operator-focused Linux fleet this is a side note, but it lowers the barrier for mixed teams, including people who restore on workstations.

Concurrency and shared repositories

This is where the three tools diverge most, and it deserves attention before you design a fleet.

  • Borg holds an exclusive lock on the repository while creating or deleting archives, so simultaneous backups can fail, and its FAQ advises against sharing a repository between clients for security reasons. One repository per client is the safe pattern, at the cost of cross-client deduplication.
  • Restic distinguishes exclusive and non-exclusive locks. Backups take a non-exclusive lock and can run in parallel against one repository, while pruning needs an exclusive lock. Locks are files in the repository, treated as stale after 30 minutes.
  • Kopia is also used with several machines connecting to one repository; its GUI exposes username and hostname options for this. The pages we read did not explain how sharing or locking works, so test concurrent runs against your own storage before committing to a design.

For any shared repository, remember the threat model. Every client that can write can usually also delete, unless you add object lock or a gatekeeping server. The site’s guide to hardening backup access goes into why that matters.

Version and maintenance notes

At the time of writing the latest Kopia release shown on the project’s GitHub releases page is v0.23.1, dated June 16, 2026. Its notes describe a fix for a rare race condition that could lead to data loss, and recommend upgrading. That is a useful reminder: Kopia is still a 0.x project, and it is worth watching release notes and testing upgrades on a non-production repository first. For Borg, the stable documentation we read was written for Borg 1.4.5 and refers to Borg 2.0 as a coming major version with an incompatible repository change and a borg transfer migration path.

You will also meet Kopia without choosing it. Velero, the Kubernetes backup tool, uses Kopia as its built-in file-system data mover, so a Velero repository is a Kopia repository. See our look at what Velero does and does not protect.

Choosing between them

PriorityLean toward
Declarative retention and per-folder policy, with a GUI availableKopia
Long track record, simple scripting, wide backend support, parallel backups to one repositoryrestic
Single-writer repositories with strong local tuning, one repository per clientBorg

All three are credible. The practical test is the same for each: initialize a repository on the storage you will really use, back up a representative data set, run a second client against it if you plan to share, break a pack on purpose, and time a full restore. Whichever one survives that exercise with a restore you can explain to a colleague is the right one for your environment.