Normally, all changes that are logged as events in the binary log are also replicated to all slaves (though still subject to filtering by --replicate-do-xxx, --replicate-ignore-xxx, and similar options). However, sometimes it may be desirable to have certain events be logged into the binlog, but not be replicated to all or a subset of slaves, where the distinction between events that should be replicated or not is under the control of the application making the changes.

This could be useful if an application does some replication external to the server outside of the built-in replication, or if it has some data that should not be replicated for whatever reason.

This is possible with two new system variables introduced in MariaDB 5.5:

Master session variable: @@skip_replication

When the skip_replication variable is set to true, changes are logged into the binary log with the flag @@skip_replication set. Such events will not be replicated by slaves that run with --replicate-events-marked-for-skip set different from its default of REPLICATE.

The skip_replication option only has effect if binary logging is enabled and sql_log_bin is true.

Attempting to change @@skip_replication in the middle of a transaction will fail; this is to avoid getting half of a transaction replicated while the other half is not replicated. Be sure to end any current transaction with COMMIT/ROLLBACK before changing the variable.

Slave option: --replicate-events-marked-for-skip

The replicate_events_marked_for_skip option tells the slave whether to replicate events that are marked with the @@skip_replication flag. Default is REPLICATE, to ensure that all changes are replicated to the slave. If set to FILTER_ON_SLAVE, events so marked will be skipped on the slave and not replicated. If set to FILTER_ON_MASTER, the filtering will be done on the master, saving on network bandwidth as the events will not be received by the slave at all.

When events are filtered due to @@skip_replication, the filtering happens on the master side; in other words, the event is never sent to the slave. If many events are filtered like this, a slave can sit a long time without receiving any events from the master. This is not a problem in itself, but must be kept in mind when inquiring on the slave about events that are filtered. For example START SLAVE UNTIL <some position> will stop when the first event that is not filtered is encountered at the given position or beyond. If the event at the given position is filtered, then the slave thread will only stop when the next non-filtered event is encountered. In effect, if an event is filtered, to the slave it appears that it was never written to the binlog on the master.

Note that when events are filtered for a slave, the data in the database will be different on the slave and on the master. It is the responsibility of the application to replicate the data outside of the built-in replication or otherwise ensure consistency of operation. If this is not done, it is possible for replication to encounter, for example, UNIQUE contraint violations or other problems which will cause replication to stop and require manual intervention to fix.

The session variable @@skip_replication can be changed without requiring special privileges. This makes it possible for normal applications to control it without requiring SUPER privileges. But it must be kept in mind when using slaves with --replicate-events-marked-for-skip set different from REPLICATE, as it allows any connection to do changes that are not replicated.

skip_replication and sql_log_bin

@@sql_log_bin and @@skip_replication are somewhat related, as they can both be used to prevent a change on the master from being replicated to the slave. The difference is that with @@skip_replication, changes are still written into the binlog, and replication of the events is only skipped on slaves that explicitly are configured to do so, with --replicate-events-marked-for-skip different from REPLICATE. With @@sql_log_bin, events are not logged into the binlog, and so are not replicated by any slave.

skip_replication and the binlog

When events in the binlog are marked with the @@skip_replication flag, the flag will be preserved if the events are dumped by the mysqlbinlog program and re-applied against a server with the mysql client program. Similarly, the BINLOG statement will preserve the flag from the event being replayed. And a slave which runs with --log-slave-updates and does not filter events (--replicate-events-marked-for-skip=REPLICATE) will also preserve the flag in the events logged into the binlog on the slave.

See also

• Using SQL_SLAVE_SKIP_COUNTER - How to skip a number of events on the slave