What to keep
- Application PostgreSQL and Hub data, including separately managed Hub databases.
- The SpiceDB datastore, plus any separate SAML database.
- Object storage, durable queue state and any other state changed by the release.
- Secrets and encryption keys, configuration, Helm values or Compose files, and exact image references.
Before an upgrade
- Stop traffic and all writers using your deployment procedure.
- Take coordinated snapshots after the final accepted writes. Include storage and separate datastores.
- Verify backups completed and record their recovery point with the current deployment configuration.
- Application database only
- Application and bundled SpiceDB (v6)
0 and verify each backup before continuing. Use your provider’s snapshot procedure for
managed databases; verify consistency and retention. Include the remaining state listed above.
Restore an instance
1
Keep maintenance active
Stop writers, schedules and reconciliation. Preserve the failed state privately for diagnosis.
Record whether any writes were accepted after the backup; restoring loses those writes.
2
Restore coordinated state
Restore databases and object storage to the same recovery point using your rehearsed datastore
procedure. Restore matching configuration, secrets and images. Avoid overwriting the only copy
of your current data while investigating.
3
Verify permissions and pending work
For a SpiceDB-backed release, check its restored schema and authorization graph with the
release-matched operations tools.
Check outbox state, allow/deny decisions and revocations before opening traffic.
Reconcile pending jobs without blindly replaying external side effects.
4
Accept the restored deployment
Run the completion checks, restore controllers
against the recovered configuration and reopen traffic. Record any lost or replayed work.
helm rollback and an older container image do not restore database state. If the release changed
stored data incompatibly, restore the backup before starting the earlier version.
To copy selected surveys rather than recover an instance, use survey transfer.