vitess-gh/doc/Reparenting.md

Reparenting is the process of changing a shard's master tablet from one host to another or changing a slave tablet to have a different master. Reparenting can be initiated manually or it can occur automatically in response to particular database conditions. As examples, you might reparent a shard or tablet during a maintenance exercise or automatically trigger reparenting when a master tablet dies.

This document explains the types of reparenting that Vitess supports:

• Active reparenting occurs when the Vitess toolchain manages the entire reparenting process.
• External reparenting occurs when another tool handles the reparenting process, and the Vitess toolchain just updates its topology server, replication graph, and serving graph to accurately reflect master-slave relationships.

Note: The InitShardMaster command defines the initial parenting relationships within a shard. That command makes the specified tablet the master and makes the other tablets in the shard slaves that replicate from that master.

MySQL requirements

Vitess supports MySQL 5.6, MySQL 5.7 and MariaDB implementations.

GTIDs

Vitess requires the use of global transaction identifiers (GTIDs) for its operations:

• During active reparenting, Vitess uses GTIDs to initialize the replication process and then depends on the GTID stream to be correct when reparenting. (During external reparenting, Vitess assumes the external tool manages the replication process.)
• During resharding, Vitess uses GTIDs for filtered replication, the process by which source tablet data is transferred to the proper destination tablets.

Semisynchronous replication

Vitess does not depend on semisynchronous replication but does work if it is implemented. Larger Vitess deployments typically do implement semisynchronous replication.

Active Reparenting

You can use the following vtctl commands to perform reparenting operations:

• PlannedReparentShard
• EmergencyReparentShard

Both commands lock the shard for write operations. The two commands cannot run in parallel, nor can either command run in parallel with the InitShardMaster command.

The two commands are both dependent on the global topology server being available, and they both insert rows in the topology server's _vt.reparent_journal table. As such, you can review your database's reparenting history by inspecting that table.

PlannedReparentShard: Planned reparenting

The PlannedReparentShard command reparents a healthy master tablet to a new master. The current and new master must both be up and running.

This command performs the following actions:

1. Puts the current master tablet in read-only mode.
2. Shuts down the current master's query service, which is the part of the system that handles user SQL queries. At this point, Vitess does not handle any user SQL queries until the new master is configured and can be used a few seconds later.
3. Retrieves the current master's replication position.
4. Instructs the master-elect tablet to wait for replication data and then begin functioning as the new master after that data is fully transferred.
5. Ensures replication is functioning properly via the following steps:
   1. On the master-elect tablet, insert an entry in a test table and then update the global Shard object's MasterAlias record.
   2. In parallel on each slave, including the old master, set the new master and wait for the test entry to replicate to the slave tablet. (Slave tablets that had not been replicating before the command was called are left in their current state and do not start replication after the reparenting process.)
   3. Start replication on the old master tablet so it catches up to the new master.

In this scenario, the old master's tablet type transitions to spare. If health checking is enabled on the old master, it will likely rejoin the cluster as a replica on the next health check. To enable health checking, set the target_tablet_type parameter when starting a tablet. That parameter indicates what type of tablet that tablet tries to be when healthy. When it is not healthy, the tablet type changes to spare.

EmergencyReparentShard: Emergency reparenting

The EmergencyReparentShard command is used to force a reparent to a new master when the current master is unavailable. The command assumes that data cannot be retrieved from the current master because it is dead or not working properly.

As such, this command does not rely on the current master at all to replicate data to the new master. Instead, it makes sure that the master-elect is the most advanced in replication within all of the available slaves.

Important: Before calling this command, you must first identify the slave with the most advanced replication position as that slave must be designated as the new master. You can use the vtctl ShardReplicationPositions command to determine the current replication positions of a shard's slaves.

This command performs the following actions:

1. Determines the current replication position on all of the slave tablets and confirms that the master-elect tablet has the most advanced replication position.
2. Promotes the master-elect tablet to be the new master. In addition to changing its tablet type to master, the master-elect performs any other changes that might be required for its new state.
3. Ensures replication is functioning properly via the following steps:
   1. On the master-elect tablet, Vitess inserts an entry in a test table and then updates the MasterAlias record of the global Shard object.
   2. In parallel on each slave, excluding the old master, Vitess sets the master and waits for the test entry to replicate to the slave tablet. (Slave tablets that had not been replicating before the command was called are left in their current state and do not start replication after the reparenting process.)

External Reparenting

External reparenting occurs when another tool handles the process of changing a shard's master tablet. After that occurs, the tool needs to call the vtctl TabletExternallyReparented command to ensure that the topology server, replication graph, and serving graph are updated accordingly.

That command performs the following operations:

1. Locks the shard in the global topology server.
2. Reads the Shard object from the global topology server.
3. Reads all of the tablets in the replication graph for the shard. Vitess does allow partial reads in this step, which means that Vitess will proceed even if a data center is down as long as the data center containing the new master is available.
4. Ensures that the new master's state is updated correctly and that the new master is not a MySQL slave of another server. It runs the MySQL show slave status command, ultimately aiming to confirm that the MySQL reset slave command already executed on the tablet.
5. Updates, for each slave, the topology server record and replication graph to reflect the new master. If the old master does not return successfully in this step, Vitess changes its tablet type to spare to ensure that it does not interfere with ongoing operations.
6. Updates the Shard object to specify the new master.

The TabletExternallyReparented command fails in the following cases:

• The global topology server is not available for locking and modification. In that case, the operation fails completely.

Active reparenting might be a dangerous practice in any system that depends on external reparents. You can disable active reparents by starting vtctld with the --disable_active_reparents flag set to true. (You cannot set the flag after vtctld is started.)

Reparenting And Serving Graph

During the reparenting process, Vitess shuffles servers such that servers might be demoted, or promoted. The serving graph should reflect the latest state of the service.

A tablet can be orphaned after a reparenting if it is unavailable when the reparent operation is running but then recovers later on. In that case, you can manually reset the tablet's master to the current shard master using the vtctl ReparentTablet command. You can then restart replication on the tablet if it was stopped by calling the vtctl StartSlave command.