Restore a database

Build a new database from a database's backups, to the latest moment, to a second in the recovery window, or to the end of one backup.

Last updated 18 November 2026View as Markdown

A restore builds a new database from the source's backups; the source is untouched. A deleted database whose backups are kept restores the same way. Replace ewdb-0123456789ab with the source database's id and orders-restored with the new database's name.

Important

The new database is billed like any other from the moment its machines are built.

Prerequisites

  • Permission for rds:RestoreDBInstance on the source database. When the new disks use a customer key (the source's, or the one you name), also kms:Encrypt on that key.
  • At least one completed backup. Otherwise the request is refused with 409 "No backup of this database has completed, so there is nothing to restore from."
  • For a time target, a moment inside the recovery window the backup list reports. Earlier than the window is refused with 400 "Nothing was kept for …", later than now with 400 "… is in the future". When only backups kept until deleted are left, only The end of one backup is accepted (409).
  • A new name: the source's own name is refused with 400 "A restore builds a new database; give it a new name.", and a name in use with 409 "A database named "…" already exists".
  • In fi1, the organization setting Allow EU locations outside Germany on at that moment (otherwise 403).
  • Room in the DB instances quota: the new database counts.
  • In the console, also rds:DescribeDBInstances to open the source database's page (without it the page is not found).

Restore

  1. Open the source database, then choose the Backups tab.
  2. Choose Restore…. The Restore to a point in time dialog opens.
  3. For New database name, enter orders-restored.
  4. (Optional) Select As a three-node cluster (the machine type needs 8 GiB or more). It is cleared by default, even for a cluster source.
  5. (Optional) Select Reachable from the internet. It is cleared by default, so a copy of a public database is never public without someone choosing it (the API instead takes the source's setting when you omit it).
  6. Under the recovery target, choose The latest moment available, A specific time (UTC) (and enter it, to the second), or The end of one backup (and choose the backup).
  7. Choose Request restore.

The dialog says what the new database is built as and what it costs a month. The request appears under Restore requests with a link to the new database.

recoveryTarget is latest (everything the repository holds), time (a moment inside the recovery window) or label with recoveryTargetLabel (the end of one backup). targetIdentifier names the new database; instanceClass defaults to the source's. publiclyAccessible decides whether the copy answers on the internet; omitted, it takes the source's setting. kmsKeyId names the key for the new disks; omitted, it takes the source's, and null uses the platform key. topology is single unless you say "cluster", whatever the source was.

Granularity. A time target can be any instant in the window, to the millisecond; the console picks to the second. Recovery replays every transaction committed up to and including that instant. A database archives its write-ahead log at least every five minutes, so the last few minutes before a request may not be in the repository yet. Before reading the backups, a time or latest restore therefore asks the source to archive what it holds (Activity: "Archived the source's current WAL"), which makes everything up to the request recoverable. If the source cannot be reached (stopped, or its machines are gone), recovery ends at the newest archived log, up to five minutes before the request. Activity warns, and a time target inside that gap fails rather than restoring to a different moment. A deleted database kept with its backups is restored up to its final backup.

Cancel a restore

While a restore is still queued (nothing has read it yet), you can withdraw it. It needs rds:RestoreDBInstance on the source database, the same grant that created the request, and the decision is recorded in the audit log.

  1. On the source database's Backups tab, under Restore requests, choose Cancel.

The job moves to cancelled, and the placeholder database the request had already created is removed with it, so the name it was holding is free again. Nothing is removed if the target's machines have already been created: a machine somebody is paying for is deleted as a database, not dropped as a row. Once the job is restoring, the request is refused with 409 "This restore has already started. It cannot be cancelled; delete the database it is building instead." Delete the new database instead (DELETE /rds/api/instances/{targetId}).

Result

The call returns 202 with the restore job and the target, which appears in the list as creating. The target has the source's engine version, storage size, master username, database name, network and retention, and the source's parameter changes, TLS-only setting and logical replication. Its first machine replaces its data directory with the chosen backup and recovers to the target before PostgreSQL first starts. On a cluster, the other two members then copy the restored database from it. The source's master password opens the target, filed in Secrets for the copy; rotating one does not change the other.

The job on GET /rds/api/instances/{id}/restore-jobs moves through queued, restoring and completed when the target is available, or failed with the reason, which is also the last Activity line on both databases.

Did this page help you?