Backup and recovery
Preserve a recoverable copy of the complete database and the deployment context
needed to open it. The current platform does not provide the previously described
/admin/snapshot or /admin/snapshots HTTP endpoints. Use a qualified operator
procedure for the deployed version.
Create an offline backup#
- Coordinate the maintenance window and stop admitting new work. Account for every client and direct writer, including administrators and background jobs.
- Stop the server and verify its exit status and logs. For versions containing the new lifecycle feature, follow graceful shutdown. A deadline or force-kill is not a successful drain.
- Confirm no process can mutate the data volume. Copy or snapshot the complete data directory and its required deployment configuration. Preserve ownership, permissions and all database subdirectories. Do not copy selected live RocksDB files or assume that separate copies of active stores form one consistent set.
- Record the deployed image digest, schema/storage versions, snapshot identity, capture time and integrity manifest. Keep the retained source unchanged.
- Restore a copy to an isolated directory or volume and open it with the exact compatible binary. Verify authorized inventories, revisions, receipts and representative reads before considering the backup usable.
Cloud volume snapshot consistency and freeze/unfreeze handling belong to the operator procedure. Snapshot completion alone does not verify restore semantics. This guide does not introduce a cloud backup service or an online snapshot API.
Restore or roll back#
Restore into a separate location first. Validate identity and authorization, retained memory, graph state and dependent learning evidence. Confirm the intended service version and actual data mounts before routing traffic to the restored instance. Preserve the failed or newer data for reconciliation.
A backup from before an upgrade may omit later acknowledged writes. Restoring it over the current directory can discard those writes; reconcile them before any cutover. Never open a migrated storage format using an older incompatible binary. Use the release's documented downgrade constraints and a verified pre-upgrade copy rather than changing only the container image.
For an interrupted shutdown, inspect the recovered database and reconcile unknown request outcomes using the original operation identities where supported. Do not infer that all interrupted requests failed or that every returned response was delivered to its client.
Set and measure recovery objectives#
The company owns retention, recovery-point and recovery-time objectives. Record measured restore duration, the actual snapshot boundary, any unreconciled writes, and validation failures. No fixed recovery-time or recovery-point guarantee is established by the examples or by a successful health check.
Rehearse restoration with realistic data and permissions. Protect backup access as database access, and include credentials/configuration needed for recovery in the approved private operational storage. Keep deployment secrets out of source repositories and user-facing evidence reports.