This task is about implementing a way for a MariaDB replication slave to
replicate from more than one master simultaneously.
Each master is handled by a specific replicator instance in the slave
server. Each replicator instance consists of separate I/O thread, SQL thread,
and associated state and configuration. In efffect, several replication slaves
are running at the same time, each replicating from a separate master, but all
replicating into a common data store (typically, but not necessarily to
separate databases/tables).
A replicator instance is identified with a user-chosen name used in
replication SQL statements such as CHANGE MASTER TO ...
This name is also included in file names to distinguish the files that keep
the replication state (relay logs, master.info, relay-log.info). This way,
each separate instance can be configured separately, but otherwise the same
way as existing single-source replication.
In order to remain backwards-compatible with existing third-party scripts
etc., the replicator instance name is made optional in all existing
replication statements. If it is omitted, the name "default" is used, and for
this particular name, master.info and the other files retain their old names
to allow seamless upgrades of slaves.
In this worklog, there is no extra conflict resolution proposed. The effect of
updates from one master on replication from another master will be the same as
the effect of direct user SQL queries on the slave server, ie. it is the
responsibility of the user/DBA to ensure no conflicts occur. If a conflict
causes replication of an event to fail (with duplicate key violation eg.), the
corresponding slave SQL thread will stop, requiring manual intervention to fix
and start.
An easy and typical way to avoid conflicts would be to eg. to use separate
databases for each master->slave replication channel. RBR idempotent slave
application can also be used to help resolve conflicts, for example.
Here is a preliminary list of things that need to be changed/extended to
handle multi-source:
File names
----------
These files used to store replication state must be extended to include the
replicator instance name. For "default", as a special case, the old name
should be used for backwards compatibility:
These statements need to be extended to take an optional replicator instance
name. If omitted, "default" is used:
CHANGE MASTER TO
LOAD DATA FROM MASTER
LOAD TABLE xxx FROM MASTER
MASTER_POS_WAIT(file, position, timeout)
RESET SLAVE
SHOW SLAVE STATUS
START SLAVE
START SLAVE UNTIL ...
STOP SLAVE
System variables
----------------
These system variables need to handle multiple replicator instances. Some can
remain global, and apply to all instances. Some will need to be per-instance,
and can probably use structured system variables, same was as for multiple key
caches:
The deprecated options master_host, master_port, master_user, master_password,
and master_ssl_* should not be supported for multi source replication.
Low-Level Design
A main part of this worklog will be to modify the server code so that all
slave objects and data can have multiple instances, not use global variables,
etc. And so that all user-visible interfaces (SQL statements, system
variables, status variables, files) are extended to support multiple
replicator instances.
Another major part is to carefully develop the feature and not the least test
it for full backwards compatibility in the case where only a single, default
master is used. If this is not done correctly, a lot of users will get
problems when slaves are upgraded and their monitoring scripts, management
scripts, or replication state breaks.
Have not been able to repeat the problem with memory allocation until very late last night. Will fix tomorrow.
The problem with sql_slave_skip_counter was that this variable is not yet multi-source aware.
Have fixed the following issues:
"Slave_IO_state and Connection_name still seem to be switched in SHOW FULL SLAVE STATUS"
--sync_with_master 0, 'master1' doesn't work
Made sql_slave_skip_counter
It's intentionally that RESET SLAVE removes the master configuration (this comes from the original patch).
The reason this method was used, is probably that there was no other logical way to remove a connection.
This does however create the problem of how to be able to remove the relay logs for a named connection.
The suggestions I have regarding this are:
Let RESET SLAVE remove relay logs and the connection but FLUSH SLAVE would only remove relay logs.
Add a DROP SLAVE 'connection_name' command.
Michael Widenius
added a comment - Have not been able to repeat the problem with memory allocation until very late last night. Will fix tomorrow.
The problem with sql_slave_skip_counter was that this variable is not yet multi-source aware.
Have fixed the following issues:
"Slave_IO_state and Connection_name still seem to be switched in SHOW FULL SLAVE STATUS"
--sync_with_master 0, 'master1' doesn't work
Made sql_slave_skip_counter
It's intentionally that RESET SLAVE removes the master configuration (this comes from the original patch).
The reason this method was used, is probably that there was no other logical way to remove a connection.
This does however create the problem of how to be able to remove the relay logs for a named connection.
The suggestions I have regarding this are:
Let RESET SLAVE remove relay logs and the connection but FLUSH SLAVE would only remove relay logs.
Add a DROP SLAVE 'connection_name' command.
Cannot we still use "RESET SLAVE 'name' ALL" to remove everything and "RESET SLAVE 'name'" to only remove the logs and position? It would be consistent with what we have for single-source: now 'RESET SLAVE ALL' removes the entire configuration, while 'RESET SLAVE' only removes logs and the position.
For the memory problem, I'll try to create a shorter test case, to reduce the amount of noise in the trace. The problem itself is reproducible on my machines and on perro, both with the plain cmake build and with BUILD/compile-pentium-debug-max. So if it doesn't happen on your boxes, maybe you could try there. I can set up the environment so you would only have to run the test.
Elena Stepanova
added a comment - Cannot we still use "RESET SLAVE 'name' ALL" to remove everything and "RESET SLAVE 'name'" to only remove the logs and position? It would be consistent with what we have for single-source: now 'RESET SLAVE ALL' removes the entire configuration, while 'RESET SLAVE' only removes logs and the position.
For the memory problem, I'll try to create a shorter test case, to reduce the amount of noise in the trace. The problem itself is reproducible on my machines and on perro, both with the plain cmake build and with BUILD/compile-pentium-debug-max. So if it doesn't happen on your boxes, maybe you could try there. I can set up the environment so you would only have to run the test.
With the new usage of max_relay_log_size we are getting extra warnings in the standard test suite (as below). Should we add the warnings to the result files, or are you planning to change the algorithm somehow?
Test lock wait timeout and purged relay logs ***
SET @my_max_relay_log_size= @@global.max_relay_log_size;
SET global max_relay_log_size=0;
+Warnings:
+Warning 1292 Truncated incorrect max_relay_log_size value: '0'
include/stop_slave.inc
DELETE FROM t2;
CHANGE MASTER TO MASTER_LOG_POS=<master_pos_begin>;
mysqltest: Result length mismatch
Elena Stepanova
added a comment - With the new usage of max_relay_log_size we are getting extra warnings in the standard test suite (as below). Should we add the warnings to the result files, or are you planning to change the algorithm somehow?
CURRENT_TEST: rpl.rpl_deadlock_innodb
— mysql-test/suite/rpl/r/rpl_deadlock_innodb.result 2012-09-28 03:23:07.482695000 +0400
+++ mysql-test/suite/rpl/r/rpl_deadlock_innodb.reject 2012-10-01 22:52:05.107989763 +0400
@@ -76,6 +76,8 @@
Test lock wait timeout and purged relay logs ***
SET @my_max_relay_log_size= @@global.max_relay_log_size;
SET global max_relay_log_size=0;
+Warnings:
+Warning 1292 Truncated incorrect max_relay_log_size value: '0'
include/stop_slave.inc
DELETE FROM t2;
CHANGE MASTER TO MASTER_LOG_POS=<master_pos_begin>;
mysqltest: Result length mismatch
Task
Major
Have not been able to repeat the problem with memory allocation until very late last night. Will fix tomorrow.
The problem with sql_slave_skip_counter was that this variable is not yet multi-source aware.
Have fixed the following issues:
It's intentionally that RESET SLAVE removes the master configuration (this comes from the original patch).
The reason this method was used, is probably that there was no other logical way to remove a connection.
This does however create the problem of how to be able to remove the relay logs for a named connection.
The suggestions I have regarding this are: