There are a number of techniques when using MySQL Replication with Global Transaction Identifiers (GTIDs) for provisioning a new slave which can then be used for scaleout, being promoted to master as necessary for failover. This section describes the following techniques:

Global transaction identifiers were added to MySQL Replication for the purpose of simplifying in general management of the replication data flow and of failover activities in particular. Each identifier uniquely identifies a set of binary log events that together make up a transaction. GTIDs play a key role in applying changes to the database: the server automatically skips any transaction having an identifier which the server recognizes as one that it has processed before. This behavior is critical for automatic replication positioning and correct failover.

The mapping between identifiers and sets of events comprising a given transaction is captured in the binary log. This poses some challenges when provisioning a new server with data from another existing server. To reproduce the identifier set on the new server, it is necessary to copy the identifiers from the old server to the new one, and to preserve the relationship between the identifiers and the actual events. This is neccessary for restoring a slave that is immediately available as a candidate to become a new master on failover or switchover.

Simple replication.  The easiest way to reproduce all identifiers and transactions on a new server is to make the new server into the slave of a master that has the entire execution history, and enable global transaction identifiers on both servers. See Section 16.1.3.2, “Setting Up Replication Using GTIDs”, for more information.

Once replication is started, the new server copies the entire binary log from the master and thus obtains all information about all GTIDs.

This method is simple and effective, but requires the slave to read the binary log from the master; it can sometimes take a comparatively long time for the new slave to catch up with the master, so this method is not suitable for fast failover or restoring from backup. This section explains how to avoid fetching all of the execution history from the master by copying binary log files to the new server.

Copying data and transactions to the slave.  Playing back the entire transaction history can be time-consuming, and represents a major bottleneck when setting up a new replication slave. To eliminate this requirement, a snapshot of the data set, the binary logs and the global transaction information the master contains is imported to the slave. The binary log is played back, after which replication can be started, allowing the slave to become current with any remaining transactions.

There are several variants of this method, the difference being in the manner in which data dumps and transactions from binary logs are transfered to the slave, as outlined here:

See also Section 4.6.7.3, “Using mysqlbinlog to Back Up Binary Log Files”.

This method has the advantage that a new server is available almost immediately; only those transactions that were committed while the snapshot or dump file was being replayed still need to be obtained from the existing master. This means that the slave's availability is not instantanteous—but only a relatively short amount of time should be required for the slave to catch up with these few remaining transactions.

Copying over binary logs to the target server in advance is usually faster than reading the entire transaction execution history from the master in real time. However, it may not always be feasible to move these files to the target when required, due to size or other considerations. The two remaining methods for provisioning a new slave discussed in this section use other means to transfer information about transactions to the new slave.

Injecting empty transactions.  The master's global gtid_executed variable contains the set of all transactions executed on the master. Rather than copy the binary logs when taking a snapshot to provision a new server, you can instead note the content of gtid_executed on the server from which the snapshot was taken. Before adding the new server to the replication chain, simply commit an empty transaction on the new server for each transaction identifier contained in the master's gtid_executed, like this:

SET GTID_NEXT='aaa-bbb-ccc-ddd:N';

BEGIN;
COMMIT;

SET GTID_NEXT='AUTOMATIC';

Once all transaction identifiers have been reinstated in this way using empty transactions, you must flush and purge the slave's binary logs, as shown here, where N is the nonzero suffix of the current binary log file name:

FLUSH LOGS;
PURGE BINARY LOGS TO 'master-bin.00000N';

You should do this to prevent this server from flooding the replication stream with false transactions in the event that it is later promoted to master. (The FLUSH LOGS statement forces the creation of a new binary log file; PURGE BINARY LOGS purges the empty transactions, but retains their identifiers.)

This method creates a server that is essentially a snapshot, but in time is able to become a master as its binary log history converges with that of the replication stream (that is, as it catches up with the master or masters). This outcome is similar in effect to that obtained using the remaining provisioning method, which we discuss in the next few paragraphs.

Excluding transactions with gtid_purged.  The master's global gtid_purged variable contains the set of all transactions that have been purged from the master's binary log. As with the method discussed previously (see Injecting empty transactions), you can record the value of gtid_executed on the server from which the snapshot was taken (in place of copying the binary logs to the new server). Unlike the previous method, there is no need to commit empty transactions (or to issue PURGE BINARY LOGS); instead, you can set gtid_purged on the slave directly, based on the value of gtid_executed on the server from which the backup or snapshot was taken.

As with the method using empty transactions, this method creates a server that is functionally a snapshot, but in time is able to become a master as its binary log history converges with that of the replication master or group.

Restoring GTID mode slaves.  When restoring a slave in a GTID based replication setup that has encountered an error, injecting an empty transaction may not solve the problem because an event does not have a GTID.

Use mysqlbinlog to find the next transaction, which is probably the first transaction in the next log file after the event. Copy everything up to the COMMIT for that transaction, being sure to include the SET @@SESSION.GTID_NEXT. Even if you are not using row-based replication, you can still run binary log row events in the command line client.

Stop the slave and run the transaction you copied. The mysqlbinlog output sets the delimiter to /*!*/;, so set it back:

mysql> DELIMITER ;

Restart replication from the correct position automatically:

mysql> SET GTID_NEXT=automatic;
mysql> RESET SLAVE;
mysql> START SLAVE;