Switchover with slave lag breaks replication about mariadb-tools HOT 9 CLOSED

That's the default behavior in latest version.

$ ./mariadb-repmgr -help
Usage of ./mariadb-repmgr:
-maxdelay=0: Maximum replication delay before initiating failover

Use -maxdelay with the value of your choice.

from mariadb-tools.

Sorry, I wasn't clear enough.
The problem occurs even if this parameter is set correctly (greater than 0)

from mariadb-tools.

OK, noted. I'll try to reproduce it.

from mariadb-tools.

OK, I suppose that happens because you use -gtidcheck=true, and that option is misdocumented and misleading (and honestly, should maybe be removed)
If you use that, failover will never happen if slaves have a different Current_GTID position from the master.
However, if you don't use it, and slaves are not beyond maxdelay, the utility will pick up the most up to date slave and will wait for GTID to synchronize before switching.
So to be clear, -gtidcheck=true is not a means to sync the slaves with the master. It's just a hardcore option to make sure that one slave is 100% in sync at switchover time.

from mariadb-tools.

Ok, understood. I found the name of this option really comfortable ;-)
Thanks for the explanations.

from mariadb-tools.

Now, a switch over with lag on all slaves result with this error on all slaves:
"Got fatal error 1236 from master when reading data from binary log: 'Error: connecting slave requested to start from GTID 1-180533626-14016589, which is not in the master's binlog"

from mariadb-tools.

Issue confirmed.

from mariadb-tools.

The reason for that issue is that switchover may happen before all transactions have completed on demoted master. It causes inconsistency in the replication setup as the new master will be behind some slaves. Looking at that I found other issues, because closing existing transactions is not handled yet. Expect some form of bug fix soon.

from mariadb-tools.

repmgr now correctly handles running threads on demoted master.
There is a configurable killswitch: -wait-kill: number of milliseconds to wait before killing connections on demoted master, if write threads are present. Reads or sleeping connections are killed no matter what.

from mariadb-tools.