Skip to content

Verify recovery across application boundaries

A recovery claim requires the data, identities, credentials and services needed by its named application scenario. Successful backup, database restore and Storage-object comparison are component facts. They do not automatically establish usable authentication, decrypted source access, application startup or a recovered business outcome.

The estate owns backup implementation. Use its generation-bound inventory and evidence through shared ownership, rather than creating a second backup system. This procedure describes the acceptance boundary; it does not authorize a shared restore.

Agree an explicitly authorized isolated target and scenario with the infrastructure owner. Prevent outbound provider effects before starting restored application processes. Identify the exact snapshot/generation, component versions, catalogue and manifests. Keep protected credentials in the approved recovery channel, separate from ordinary evidence.

The current owner document records a 2026-09-30 database and complete Storage restore from the same exact B2 snapshot/generation. Counts and object hashes passed in those component checks. They did not establish combined application startup, Auth/credential recovery, mail decryption, measured RTO or recurring recovery success.

  1. Inventory what the scenario needs: database records and grants, Storage metadata and bytes, Auth identities/access, runtime configuration, component code and any encryption identity. State exclusions before executing.
  2. Verify backup identity, timestamp, component coverage and hashes. A current service exit code is not evidence that every required artifact belongs to the same recoverable generation.
  3. Restore only to the authorized isolated target using the estate procedure and the repository’s image-specific requirements. Wait for full container health where required; pg_isready alone can precede image initialization. Never overwrite the persistent shared database or import password hashes casually.
  4. Before compatibility migrations, use scripts/db/schema_catalogue.py to compare the restored schema with the generation’s recorded catalogue. Compare count-only inventories and Storage object hashes. supabase db diff cannot reliably represent all objects in this schema.
  5. Recover the necessary Auth/configuration/decryption boundary through its owning procedure. The mail-intake manifest requires database, ciphertext objects and a recovered age identity with a decryption fixture together. Its declared_not_verified status is not a successful drill.
  6. Start the compatible application with provider effects safely disabled. Verify the named actor’s access, authorized record/files and affected application behaviour. Keep synthetic destructive fixtures on their own target, never the restored customer dataset.
  7. Record recovery-point identity, start/end times, measured elapsed interval, checks, exclusions and failures without payloads or secrets. Compare actual measurements with objectives only within this scenario.
Check Establishes Still needs separate evidence
Backup metadata and hashes Named artifacts were preserved intact Restorability and generation coverage
Database catalogue/counts Compared definitions and inventory match Record semantics, object access and application outcome
Storage bytes and hashes Named object contents match Correct authorization and linked-record behaviour
Auth/configuration/decryption Required identity paths work within the drill Complete affected business scenario
Application scenario Named outcome on the recovered target Other scenarios, recurring proof and unrelated provider state

Stop at a missing generation, mismatched catalogue, absent identity or unsafe outbound path. Preserve safe diagnostics and obtain the owning correction. Do not delete a usable backup, rewrite evidence or point reset/seed tools at restored private data to obtain a pass.

RPO 60 minutes and RTO 240 minutes remain objectives associated with the open recovery outcome. Nightly backups and no established PITR do not prove the RPO objective. A component drill without combined application timing does not prove RTO. Keep those gaps explicit in limitations.

Application rollback selects a previous compatible release; it does not restore the database or reverse provider effects. Schema publication, credential recovery and evidence interpretation identify the separate owners. A synthetic recovery of one project and attachment can prove that scenario while honestly leaving payroll, mail decryption or historical reconciliation outside scope.