Skip to main content
This page covers existing local offline commands. A bundle contains one database schema and matching private identity/key material, not a complete live Sandbox snapshot or hosted infrastructure backup.

Know what the bundle can recover

A closed-Session guard is not a provider-wide orphan audit. Confirm the resources associated with the backed-up workload and retain unresolved records separately. Offline consistency also requires that no other process writes the selected database/settings while the backup is taken; stopping one launcher is not proof that every external writer is stopped.

Prepare an offline backup

Install pg_dump/pg_restore and select the correct DATABASE_URL/HEXAGON_DATA_DIR. Settle Runs, close every cloud Session with confirmed cleanup, and stop the local service. The implementation rejects running services, unclosed Sessions, and mismatched identity/decryption.
The destination must be new. Bundles contain database.dump, master.key, owner.json, optional deployment.json, hashes, and migration metadata. Private permissions are used; do not commit or publicly upload bundles.

Restore into a fresh target

Create an empty PostgreSQL database and an absent/empty data directory. Replace this illustrative connection with your dedicated target:
Restore validates the manifest, file types, and hashes, rejects a nonempty database, restores transactionally, then verifies owner identity and decryption of encrypted records. For a non-public returned schema, configure the subsequent connection search_path. Start with the target database/data directory rather than accidentally starting the old workspace. Discovery and PID files are regenerated. If failure occurs after database restoration, the target may no longer be empty. Preserve diagnostics and inspect staging/database state instead of blindly repeating or removing data.

Recover from an interrupted restore

The entire command is not one atomic transaction across PostgreSQL and the filesystem. A second attempt against a populated target is correctly refused. Do not delete that target as a generic retry step or start both old and restored deployments against the same remote resources.

Trust and consistency

Restore only trusted bundles. Hashes establish integrity against a manifest, not author trust. The database and key must match. The local command requires closed Sessions to avoid resurrecting live operations. Independently restoring older or online database copies requires additional reconciliation of later cloud resources and external writes; do not assume safe replay. Hosted external key storage, identity, TLS, and deployment configuration need their own backup process. Local restore success does not establish hosted recovery.

Verify

In an isolated target, read history, initialization, events, and artifacts; verify decryption; run a fresh task and confirm cleanup; ensure old state did not unexpectedly recreate resources. Record commands, schema, scope, and failures without publishing backup files or secrets. Implementation: offline backup/restore. Next: reconciliation, upgrades, persistence. The November 18 isolated restore drill compared all 27 then-existing tables by sorted-row hash, including Sessions, Runs, events, artifacts, connections and encrypted secrets, and subsequently ran both SDK cloud journeys from the restored deployment. These counts describe that checkpoint, not today’s schema inventory. It is local-mode evidence; no new restore was performed for this page.