ABS — backup at the moment of change
An industrial backup system with continuous synchronisation: its own storage format with deduplication and snapshots, its own cryptography, its own network core — from a file on a workstation to the store, with no third-party parts.
Backup has become someone else's infrastructure
- Western vendors left. Licences aren't renewed, there are no updates or support, and the data sits in their closed format.
- The cloud is someone else's server. The copy goes where more than its owners can read it.
- The nightly window. A copy is taken once a day; between copies the data is unprotected.
- Restores go untested. A damaged archive is discovered on the day of the disaster.
- Storage grows linearly. Without deduplication every copy is another full volume.
A copy is taken when a file changes
| Step | What happens |
|---|---|
| 1 · Event | the AWH core reports a change, rename, deletion or permission change — immediately, with no tree walk |
| 2 · Chunking | the file is cut into content-defined blocks; a block's hash is its address and checksum |
| 3 · Selection | blocks already in the store are not sent at all — neither over the network nor to disk |
| 4 · Protection | new blocks are compressed with a chosen codec and encrypted with the client's key |
| 5 · Snapshot | the store commits the state: any version can be retrieved whole or file by file |
Changing 1 % of the data costs about 3 % of traffic — an acceptance criterion of the project. The server stores only ciphertext and holds no keys at all.
File system subscription is already in the AWH core
Watching files and directories is built into the network core and uses each system's native mechanism — kqueue, epoll, io_uring, IOCP and event ports. Competitors solve this with a separate driver per OS or with periodic tree walks.
The data-loss window is minutes, not a day. The core is woken by changes, not a schedule: an idle fleet of machines is never scanned.
A smart container instead of an archive
At the base is the second version of the Crisper container: a store that behaves like a file system.
- Content addressing — identical blocks are stored once: a hundred copies of one fleet take little more than one.
- Copy-on-write index — a power cut never destroys the previous version.
- Free snapshots — every commit is a version.
- Partitions — losing one part doesn't destroy the rest and is repaired separately.
- A Merkle tree with a signed root — tampering is detected, not assumed.
- Per-block compression — ten codecs and dictionaries for millions of small files.
A store that cannot read your data
Client side
Blocks are encrypted before sending; each one's integrity is checked on decryption. Changing a password re-encrypts 32 bytes, not terabytes.
Our own cryptography
A-NTS: Kuznyechik, Magma, Streebog, GOST R 34.10-2012 signatures, plus AES-GCM, ChaCha20-Poly1305, X25519.
Signature and trail
The manifest is signed with the owner's key; who committed each version, and when, is visible.
Zero knowledge here is a consequence of the design: keys never leave the client's perimeter, and the store operator has nothing to hand over even with full access to the disks.
Damage is found in advance, files are found without a restore
- Scheduled checks name the specific damaged blocks and files.
- Extraction despite damage — damaged parts are skipped, the rest comes back whole.
- Self-repair from partition redundancy, without a second copy.
- Mount a backup as a disk — one document takes minutes, not a restore of the whole night's snapshot.
- Search by name with typos and by content.
- Metadata survives moves — POSIX rights, NTFS ACLs, links, Cyrillic names.
Where ABS beats what clients run today
| ABS | tar + scripts | restic / Kopia | Veeam / Acronis | |
|---|---|---|---|---|
| Block deduplication | yes | no | yes | yes |
| Continuous protection on FS events | yes | no | no | partly |
| Backup as a disk | yes | no | yes | yes |
| Full-text search in backups | yes | no | no | separate product |
| Repair of damaged backups | yes | no | check only | check only |
| GOST included | yes | no | no | no |
| No internet, no vendor | yes | yes | yes | licence |
restic and Kopia are honest open projects and the closest to us. The difference is industrial support, national cryptography, a desktop client and continuous protection — for us these are parts of one product.
Most of it is already written
| Part | What it provides | Status |
|---|---|---|
| AWH | networking, HTTP/1–3, QUIC, cluster, file system subscription | in development |
| A-NTS | our own TLS, GOST, classics and post-quantum | in development |
| Crisper | container, partitions, compression, encryption, search, desktop client | available |
| Smart container and the ABS service | what is being built: a 12-stage plan with acceptance criteria | planned |