On this page
flowchart LR
S["running server\n(stop writes)"] --> Q["quiesce\ncheckpoint + close"] --> C["copy data dir\n+ manifest"] --> B["backup"]
B --> I["backup-inspect\nverify manifest"]
B --> R["restore to CLEAN dir"]
R --> V["start + verify\ncounts, index, txn, write"]Diagram source · mermaidcopy included
flowchart LR
S["running server\n(stop writes)"] --> Q["quiesce\ncheckpoint + close"] --> C["copy data dir\n+ manifest"] --> B["backup"]
B --> I["backup-inspect\nverify manifest"]
B --> R["restore to CLEAN dir"]
R --> V["start + verify\ncounts, index, txn, write"]Cold full backup only: stop writes (checkpoint and close the engine, or stop the server), copy the data directory, restore to a clean directory. There is no hot backup, no incremental backup, and no point-in-time recovery in Beta 2.
Backup#
- Checkpoint: ensure the engine checkpoints (graceful shutdown checkpoints automatically), then stop the server or close all writers.
- Copy the entire data directory to the backup location. The copy must be taken while no writer is running. The
plomid-server backupcommand performs the quiesce (open, checkpoint, close) and the copy in one fenced step — see below. - The backup ends with a manifest (
PLOMID_BACKUP_MANIFEST). A backup without a manifest is an interrupted backup — discard it, never restore from it.
Offline backup commands#
The same binary ships fenced offline tools (they never start a listener). Each mutating command takes an advisory lock on its target directory and refuses to run against a live server instead of copying under a writer:
plomid-server backup --data /var/lib/plomid/data --to /backups/plomid-2026-10-14
plomid-server backup-inspect --backup /backups/plomid-2026-10-14
plomid-server restore --from /backups/plomid-2026-10-14 --to /var/lib/plomid/restoredplomid-server help-backup prints the same contract. Both --flag value and --flag=value forms work.
backup checkpoints and closes the engine before copying; restore additionally refuses non-clean destinations; backup-inspect is read-only and prints the manifest (format versions, file count, bytes, timestamp). Exit codes: 0 ok, 1 usage, 2 target live or refused, 3 operation failed (message on stderr in all cases).
Excluded by design: *.tmp staged files and the derived vector_ann cache (rebuilt from rows on reopen).
Restore#
- Restore to a clean (empty or nonexistent) directory only. Restoring over a live database is refused — this is the rollback policy, not a limitation to work around.
- Start the server against the restored directory. Recovery replays the WAL tail automatically; torn tails are truncated, corruption fails loud.
- Verify: row counts and checksums (
COUNT(*),SUM), index lookups, one transaction (BEGIN … COMMIT), and one new write followed by a second reopen.
Upgrade and rollback#
- Upgrade: cold backup on the old binary → restore to a clean directory with the new binary → verify. A version-skewed image fails with
Unsupported(never a silent downgrade) — upgrade the binary instead. - Rollback: restore the pre-upgrade cold backup to a clean directory. There is no in-place downgrade path.
Limits#
- Full copies only; large databases need proportional disk and time.
COPY TOoutput, audit logs, and TLS keys are not part of a database backup — back those up separately.