Sign InTry Free

Online Unsafe Recovery

When permanently damaged replicas cause part of data on TiKV to be unreadable and unwritable, you can use the Online Unsafe Recovery feature to perform a lossy recovery operation.

Feature description

In TiDB, the same data might be stored in multiple stores at the same time according to the replica rules defined by users. This guarantees that data is still readable and writable even if a single or a few stores are temporarily offline or damaged. However, when most or all replicas of a Region go offline during a short period of time, the Region becomes temporarily unavailable and cannot be read or written.

Suppose that multiple replicas of a data range encounter issues like permanent damage (such as disk damage), and these issues cause the stores to stay offline. In this case, this data range is temporarily unavailable. If you want the cluster back in use and also accept data rewind or data loss, in theory, you can re-form the majority of replicas by manually removing the failed replicas from the group. This allows application-layer services to read and write this data range (might be stale or empty) again.

In this case, if some stores with loss-tolerant data are permanently damaged, you can perform a lossy recovery operation by using Online Unsafe Recovery. When you use this feature, PD automatically pauses Region scheduling (including split and merge), collects the metadata of data shards from all stores, and then, under its global perspective, generates a real-time and complete recovery plan. Then, PD distributes the plan to all surviving stores to make them perform data recovery tasks. In addition, once the data recovery plan is distributed, PD periodically monitors the recovery progress and re-send the plan when necessary.

User scenarios

The Online Unsafe Recovery feature is suitable for the following scenarios:

  • The data for application services is unreadable and unwritable, because permanently damaged stores cause the stores to fail to restart.
  • You can accept data loss and want the affected data to be readable and writable.
  • You want to perform a one-stop online data recovery operation.



Before using Online Unsafe Recovery, make sure that the following requirements are met:

  • The offline stores indeed cause some pieces of data to be unavailable.
  • The offline stores cannot be automatically recovered or restarted.

Step 1. Specify the stores that cannot be recovered

To trigger automatic recovery, use PD Control to execute unsafe remove-failed-stores <store_id>[,<store_id>,...] and specify all the TiKV nodes that cannot be recovered, separated by commas.

pd-ctl -u <pd_addr> unsafe remove-failed-stores <store_id1,store_id2,...>

If the command returns Success, PD Control has successfully registered the task to PD. This only means that the request has been accepted, not that the recovery has been successfully performed. The recovery task is performed in the background. To see the recovery progress, use show.

If the command returns Failed, PD Control has failed to register the task to PD. The possible errors are as follows:

  • unsafe recovery is running: There is already an ongoing recovery task.
  • invalid input store x doesn't exist: The specified store ID does not exist.
  • invalid input store x is up and connected: The specified store with the ID is still healthy and should not be recovered.

To specify the longest allowable duration of a recovery task, use the --timeout <seconds> option. If this option is not specified, the longest duration is 5 minutes by default. When the timeout occurs, the recovery is interrupted and returns an error.

Step 2. Check the recovery progress and wait for the completion

When the above store removal command runs successfully, you can use PD Control to check the removal progress by running unsafe remove-failed-stores show.

pd-ctl -u <pd_addr> unsafe remove-failed-stores show

The recovery process has multiple possible stages:

  • collect report: The initial stage in which PD collects reports from TiKV and gets global information.
  • tombstone tiflash learner: Among the unhealthy Regions, delete the TiFlash learners that are newer than other healthy peers, to prevent such an extreme situation and the possible panic.
  • force leader for commit merge: A special stage. When there is an uncompleted commit merge, force leader is first performed on the Regions with commit merge, in case of extreme situations.
  • force leader: Forces unhealthy Regions to assign a Raft leader among the remaining healthy peers.
  • demote failed voter: Demotes the Region's failed voters to learners, and then the Regions can select a Raft leader as normal.
  • create empty region: Creates an empty Region to fill in the space in the key range. This is to resolve the case that the stores with all replicas of some Regions have been damaged.

Each of the above stages is output in the JSON format, including information, time, and a detailed recovery plan. For example:

[ { "info": "Unsafe recovery enters collect report stage: failed stores 4, 5, 6", "time": "......" }, { "info": "Unsafe recovery enters force leader stage", "time": "......", "actions": { "store 1": [ "force leader on regions: 1001, 1002" ], "store 2": [ "force leader on regions: 1003" ] } }, { "info": "Unsafe recovery enters demote failed voter stage", "time": "......", "actions": { "store 1": [ "region 1001 demotes peers { id:101 store_id:4 }, { id:102 store_id:5 }", "region 1002 demotes peers { id:103 store_id:5 }, { id:104 store_id:6 }", ], "store 2": [ "region 1003 demotes peers { id:105 store_id:4 }, { id:106 store_id:6 }", ] } }, { "info": "Collecting reports from alive stores(1/3)", "time": "......", "details": [ "Stores that have not dispatched plan: ", "Stores that have reported to PD: 4", "Stores that have not reported to PD: 5, 6", ] } ]

After PD has successfully dispatched the recovery plan, it waits for TiKV to report the execution results. As you can see in Collecting reports from alive stores, the last stage of the above output, this part of the output shows the detailed statuses of PD dispatching recovery plan and receiving reports from TiKV.

The whole recovery process takes multiple stages and one stage might be retried multiple times. Usually, the estimated duration is 3 to 10 periods of store heartbeat (one period of store heartbeat is 10 seconds by default). After the recovery is completed, the last stage in the command output shows "Unsafe recovery finished", the table IDs to which the affected Regions belong (if there is none or RawKV is used, the output does not show the table IDs), and the affected SQL meta Regions. For example:

{ "info": "Unsafe recovery finished", "time": "......", "details": [ "Affected table ids: 64, 27", "Affected meta regions: 1001", ] }

After you get the affected table IDs, you can query INFORMATION_SCHEMA.TABLES to view the affected table names.


If an error occurs during the task, the last stage in the output shows "Unsafe recovery failed" and the error message. For example:

{ "info": "Unsafe recovery failed: <error>", "time": "......" }

Step 3. Check the consistency of data and index (not required for RawKV)

After the recovery is completed, the data and index might be inconsistent. Use the SQL command ADMIN CHECK to check the data and index consistency of the affected tables


If there are inconsistent indexes, you can fix the index inconsistency by renaming the old index, creating a new index, and then dropping the old index.

  1. Rename the old index:

    ALTER TABLE table_name RENAME INDEX index_name TO index_name_lame_duck;
  2. Create a new index:

    ALTER TABLE table_name ADD INDEX index_name (column_name);
  3. Drop the old index:

    ALTER TABLE table_name DROP INDEX index_name_lame_duck;

Step 4: Remove unrecoverable stores (optional)

  • Stores deployed using TiUP
  • Stores deployed using TiDB Operator
  1. Remove the unrecoverable nodes:

    tiup cluster scale-in <cluster-name> -N <host> --force
  2. Clean up Tombstone nodes:

    tiup cluster prune <cluster-name>
  1. Delete the PersistentVolumeClaim.

    kubectl delete -n ${namespace} pvc ${pvc_name} --wait=false
  2. Delete the TiKV Pod and wait for newly created TiKV Pods to join the cluster.

    kubectl delete -n ${namespace} pod ${pod_name}

Was this page helpful?

Download PDFRequest docs changesAsk questions on Discord
One-stop & interactive experience of TiDB's capabilities WITHOUT registration.
TiDB Dedicated
TiDB Serverless
Get Demo
Get Started
© 2024 PingCAP. All Rights Reserved.
Privacy Policy.