Release v0.2.0 · soroushalinia/backupd

v0.2.0 - 2026-07-31

Hardlinks: detected during the file walk (device + inode), stored once (first path carries
the blocks, others record a link reference), and restored as real hardlinks - the link
relationship survives a restore.
Doublestar exclude patterns: gitignore-style globs (**, *, ?, {a,b}) matched against
the full relative path, the basename, and any path segment. Substring matching is gone (cache
does not match cachedir).
Streaming multipart uploads for container archives: docker and kubernetes dumps stream as
8 MiB multipart parts (encrypted in flight), no spooling to temporary files, aborted cleanly on
failure. Storages without multipart support fall back to the spool path.
backupd prune <plan>: applies the plan's retention policy on demand, then
garbage-collects orphaned blocks. --dry-run reports what would be deleted without deleting.
restore --dry-run: reads the manifest and reports each source (file count, bytes, blocks)
and whether archive objects are present, without writing anything.
Per-file upload logging: each changed file reports the bytes actually uploaded and its block
count; unchanged files log as such; plans with a rate limit announce it up front.

Prune could corrupt kept snapshots: collectBlocks ignored source-level database dump
blocks, so pruning deleted blocks still referenced by kept snapshots. Fixed, and manifest parse
errors now propagate instead of silently skipping.
Prune never deleted source archives of expired snapshots (deleteSnapshot double-prefixed
keys).
Orphaned blocks now reclaimed immediately: garbage collection runs after failed runs and
on plans without a retention policy, not only during retention pruning.
Failing capture tools now fail the run: a database dump or container capture that exits
nonzero mid-stream produced a truncated snapshot that was stored and verified as healthy; the
exit status is now checked and the run aborts (uploaded objects are cleaned up).
Stable encryption salt: the per-plan salt is loaded from the most recent manifest instead
of being regenerated, so encrypted incremental snapshots verify.

Large-file round trip (96 MiB, ~12k content-addressed blocks), engine-level traversal
rejection for malicious manifests (file paths and hardlink targets), failed runs with plain
file sources (no docker required), multipart round trip, dump-tool failure handling, hardlink
detection/restore, database dump dedup across runs.

First release.

Incremental delta backups: files are hashed once per run, unchanged files upload nothing,
changed files split into 8 KiB content-addressed blocks (uploaded only when the object does not
already exist in the bucket).
Database dumps go through the same block pipeline - an unchanged database uploads only the
manifest.
Convergent AES-256-GCM encryption for blocks and stream-encrypted archives (1 MiB frames),
Argon2id key derivation, stable per-plan salt.
Retention policies (keep-last / keep-daily / keep-weekly / keep-monthly) with automatic
pruning and orphaned-block garbage collection.
Container sources: docker volumes and kubernetes PVCs snapshotted as archives.
Symlinks and empty directories backed up; restore paths contained under the target directory.
On-demand and scheduled runs (backupd daemon, systemd export), pre/post/on-failure hooks,
per-plan rate limiting, backupd check and run --dry-run.
Failed runs clean up their partial objects; snapshot state is recorded only on success.