RESTORE

This statement performs a distributed restore from a backup archive previously produced by a BACKUP statement.

The RESTORE statement uses the same engine as the BR tool, except that the restore process is driven by TiDB itself rather than a separate BR tool. All benefits and caveats of BR also apply here. In particular, RESTORE is currently not ACID-compliant. Before running RESTORE, ensure that the following requirements are met:

• The cluster is "offline", and the current TiDB session is the only active SQL connection to access all tables being restored.
• When a full restore is being performed, the tables being restored should not already exist, because existing data might be overridden and causes inconsistency between the data and indices.
• When an incremental restore is being performed, the tables should be at the exact same state as the LAST_BACKUP timestamp when the backup is created.

Running RESTORE requires either the RESTORE_ADMIN or SUPER privilege. Additionally, both the TiDB node executing the restore and all TiKV nodes in the cluster must have read permission from the destination.

The RESTORE statement is blocking, and will finish only after the entire restore task is finished, failed, or canceled. A long-lasting connection should be prepared for running RESTORE. The task can be canceled using the KILL TIDB QUERY statement.

Only one BACKUP and RESTORE task can be executed at a time. If a BACKUP or RESTORE task is already running on the same TiDB server, the new RESTORE execution will wait until all previous tasks are done.

RESTORE can only be used with "tikv" storage engine. Using RESTORE with the "unistore" engine will fail.

Synopsis

RestoreStmt ::=
    "RESTORE" BRIETables "FROM" stringLit RestoreOption*

BRIETables ::=
    "DATABASE" ( '*' | DBName (',' DBName)* )
|   "TABLE" TableNameList

RestoreOption ::=
    "RATE_LIMIT" '='? LengthNum "MB" '/' "SECOND"
|   "CONCURRENCY" '='? LengthNum
|   "CHECKSUM" '='? Boolean
|   "SEND_CREDENTIALS_TO_TIKV" '='? Boolean

Boolean ::=
    NUM | "TRUE" | "FALSE"

Examples

Restore from backup archive

RESTORE DATABASE * FROM 'local:///mnt/backup/2020/04/';

+------------------------------+-----------+----------+---------------------+---------------------+
| Destination                  | Size      | BackupTS | Queue Time          | Execution Time      |
+------------------------------+-----------+----------+---------------------+---------------------+
| local:///mnt/backup/2020/04/ | 248665063 | 0        | 2020-04-21 17:16:55 | 2020-04-21 17:16:55 |
+------------------------------+-----------+----------+---------------------+---------------------+
1 row in set (28.961 sec)

In the example above, all data is restored from a backup archive at the local filesystem. The data is read as SST files from the /mnt/backup/2020/04/ directories distributed among all TiDB and TiKV nodes.

The first row of the result above is described as follows:

Partial restore

You can specify which databases or tables to restore. If some databases or tables are missing from the backup archive, they will be ignored, and thus RESTORE would complete without doing anything.

RESTORE DATABASE `test` FROM 'local:///mnt/backup/2020/04/';

RESTORE TABLE `test`.`sbtest01`, `test`.`sbtest02` FROM 'local:///mnt/backup/2020/04/';

External storages

BR supports restoring data from S3 or GCS:

RESTORE DATABASE * FROM 's3://example-bucket-2020/backup-05/';

The URL syntax is further explained in external storage URL.

When running on cloud environment where credentials should not be distributed, set the SEND_CREDENTIALS_TO_TIKV option to FALSE:

RESTORE DATABASE * FROM 's3://example-bucket-2020/backup-05/'
    SEND_CREDENTIALS_TO_TIKV = FALSE;

Performance fine-tuning

Use RATE_LIMIT to limit the average download speed per TiKV node to reduce network bandwidth.

By default, TiDB node would run 128 restore threads. This value can be adjusted with the CONCURRENCY option.

Before restore is completed, RESTORE would perform a checksum against the data from the archive to verify correctness. This step can be disabled with the CHECKSUM option if you are confident that this is unnecessary.

RESTORE DATABASE * FROM 's3://example-bucket-2020/backup-06/'
    RATE_LIMIT = 120 MB/SECOND
    CONCURRENCY = 64
    CHECKSUM = FALSE;

Incremental restore

There is no special syntax to perform incremental restore. TiDB will recognize whether the backup archive is full or incremental and take appropriate action. You only need to apply each incremental restore in correct order.

For instance, if a backup task is created as follows:

BACKUP DATABASE `test` TO 's3://example-bucket/full-backup'  SNAPSHOT = 413612900352000;
BACKUP DATABASE `test` TO 's3://example-bucket/inc-backup-1' SNAPSHOT = 414971854848000 LAST_BACKUP = 413612900352000;
BACKUP DATABASE `test` TO 's3://example-bucket/inc-backup-2' SNAPSHOT = 416353458585600 LAST_BACKUP = 414971854848000;

then the same order should be applied in the restore:

RESTORE DATABASE * FROM 's3://example-bucket/full-backup';
RESTORE DATABASE * FROM 's3://example-bucket/inc-backup-1';
RESTORE DATABASE * FROM 's3://example-bucket/inc-backup-2';

MySQL compatibility

This statement is a TiDB extension to MySQL syntax.

See also

• BACKUP
• SHOW RESTORES