[BUG] Switch over with mariadb-repmgr and multi GTID domains looks broken

[cpeintre]

Original state:

Master Host   Port      Binlog Position  Strict Mode
srv-02         3306 1-180533626-105,2-180533626-1          OFF

 Slave Host   Port  Binlog   Using GTID           Slave GTID   Replication Health  Delay  RO
  srv-03   3306      ON    Slave_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON
  srv-04   3306      ON  Current_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON
  srv-05   3306      ON    Slave_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON
  srv-06   3306      ON  Current_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON

Then, state after a switch over:

Master Host   Port      Binlog Position  Strict Mode
srv-03         3306       1-180533627-92          OFF

 Slave Host   Port  Binlog   Using GTID           Slave GTID   Replication Health  Delay  RO
  srv-02   3306      ON  Current_Pos      1-180533630-104           Running OK      0  ON
  srv-04   3306      ON  Current_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON
  srv-05   3306      ON    Slave_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON
  srv-06   3306      ON  Current_Pos 1-180533626-105,2-180533626-1           Running OK      0  ON

The new master and the old master are not in sync with other slaves.

State after a new transaction:

Master Host   Port      Binlog Position  Strict Mode
srv-03         3306      1-180533627-106          OFF

 Slave Host   Port  Binlog   Using GTID           Slave GTID   Replication Health  Delay  RO
  srv-02   3306      ON  Current_Pos      1-180533627-106           Running OK      0  ON
  srv-04   3306      ON  Current_Pos 1-180533627-106,2-180533626-1           Running OK      0  ON
  srv-05   3306      ON    Slave_Pos 1-180533627-106,2-180533626-1           Running OK      0  ON
  srv-06   3306      ON  Current_Pos 1-180533627-106,2-180533626-1           Running OK      0  ON

Switch over is no longer possible.

[cpeintre]

Need a new transaction on all domain id to retrieve a valid state:

 Master Host   Port      Binlog Position  Strict Mode
srv-03         3306 1-180533627-106,2-180533627-2          OFF

 Slave Host   Port  Binlog   Using GTID           Slave GTID   Replication Health  Delay  RO
  srv-02   3306      ON  Current_Pos 1-180533627-106,2-180533627-2           Running OK      0  ON
  srv-04   3306      ON  Current_Pos 1-180533627-106,2-180533627-2           Running OK      0  ON
  srv-05   3306      ON    Slave_Pos 1-180533627-106,2-180533627-2           Running OK      0  ON
  srv-06   3306      ON  Current_Pos 1-180533627-106,2-180533627-2           Running OK      0  ON

[tanji]

Yes, this is a known limitation (I know, should document them...). However I have ways to overcome them. This will be fixed in a next release.

[siddontang]

mysqlfailover tries to let the candidate sync the missing transactions from other slaves first if possible, but I still think multi source replication is complex in production environment, I like keeping replication topology simple. :-)

[tanji]

fixed in main repo