Pre-General Availability Draft: 2018-02-24
MySQL supports delayed replication such that a slave server deliberately executes transactions later than the master by at least a specified amount of time. This section describes how to configure a replication delay on a slave, and how to monitor replication delay.
In MySQL 8.0, the method of delaying replication
depends on two timestamps,
Replication Delay Timestamps). If all
servers in the replication topology are running MySQL 8.0.1 or
above, delayed replication is measured using these timestamps. If
either the immediate master or slave is not using these
timestamps, the implementation of delayed replication from MySQL
5.7 is used (see
Delayed Replication). This
section describes delayed replication between servers which are
all using these timestamps.
The default replication delay is 0 seconds. Use the
CHANGE MASTER TO MASTER_DELAY=N statement to
set the delay to
N seconds. A
transaction received from the master is not executed until at
N seconds later than its commit
on the immediate master. The delay happens per transaction (not
event as in previous MySQL versions) and the actual delay is
imposed only on
anonymous_gtid_log_event. The other events in
the transaction always follow these events without any waiting
time imposed on them.
Performance Schema table contains the
DESIRED_DELAY column which shows the delay
configured using the
MASTER_DELAY option. The
Performance Schema table contains the
REMAINING_DELAY column which shows the number
of delay seconds remaining.
Delayed replication can be used for several purposes:
To protect against user mistakes on the master. With a delay you can roll back a delayed slave to the time just before the mistake.
To test how the system behaves when there is a lag. For example, in an application, a lag might be caused by a heavy load on the slave. However, it can be difficult to generate this load level. Delayed replication can simulate the lag without having to simulate the load. It can also be used to debug conditions related to a lagging slave.
To inspect what the database looked like in the past, without having to reload a backup. For example, by configuring a slave with a delay of one week, if you then need to see what the database looked like before the last few days' worth of development, the delayed slave can be inspected.
MySQL 8.0 provides a new method for measuring delay (also referred to as replication lag) in replication topologies that depends on the following timestamps associated with the GTID of each transaction (instead of each event) written to the binary log.
original_commit_timestamp: the number of microseconds since epoch when the transaction was written (committed) to the binary log of the original master.
immediate_commit_timestamp: the number of microseconds since epoch when the transaction was written (committed) to the binary log of the immediate master.
The output of mysqlbinlog displays these
timestamps in two formats, microseconds from epoch and also
TIMESTAMP format, which is based on the user
defined timezone for better readability. For example:
#170404 10:48:05 server id 1 end_log_pos 233 CRC32 0x016ce647 GTID last_committed=0 \ sequence_number=1 original_committed_timestamp=1491299285661130 immediate_commit_timestamp=1491299285843771 # original_commit_timestamp=1491299285661130 (2017-04-04 10:48:05.661130 WEST) # immediate_commit_timestamp=1491299285843771 (2017-04-04 10:48:05.843771 WEST) /*!80001 SET @@session.original_commit_timestamp=1491299285661130*//*!*/; SET @@SESSION.GTID_NEXT= 'aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa:1'/*!*/; # at 233
As a rule, the
always the same on all replicas where the transaction is
applied. In master-slave replication, the
original_commit_timestamp of a transaction in
the (original) master’s binary log is always the same as its
immediate_commit_timestamp. In the slave’s
relay log, the
immediate_commit_timestamp of the transaction
are the same as in the master’s binary log; whereas in its own
binary log, the transaction’s
immediate_commit_timestamp corresponds to
when the slave committed the transaction.
In a Group Replication setup, when the original master is a
member of a group, the
original_commit_timestamp is generated when
the transaction is ready to be committed. In other words, when
it finished executing on the original master and its write set
is ready to be sent to all members of the group for
certification. Therefore, the same
original_commit_timestamp is replicated to
all servers (regardless of whether it is a group member or slave
replicating from a member) applying the transaction and each
stores in its binary log the local commit time using
View change events, which are exclusive to Group Replication,
are a special case. Transactions containing these events are
generated by each server but share the same GTID (so, they are
not first executed in a master and then replicated to the group,
but all members of the group execute and apply the same
transaction). Since there is no original master, these
transactions have their
original_commit_timestamp set to zero.
One of the most common ways to monitor replication delay (lag)
in previous MySQL versions was by relying on the
Seconds_Behind_Master field in the output of
SHOW SLAVE STATUS. However, this metric is
not suitable when using replication topologies more complex than
the traditional master-slave setup, such as Group Replication.
The addition of
original_commit_timestamp to MySQL 8
provides a much finer degree of information about replication
delay. The recommended method to monitor replication delay in a
topology that supports these timestamps is using the following
Performance Schema tables.
replication_connection_status: current status of the connection to the master, provides information on the last and current transaction the connection thread queued into the relay log.
replication_applier_status_by_coordinator: current status of the coordinator thread that only displays information when using a multi-threaded slave, provides information on the last transaction buffered by the coordinator thread to a worker’s queue, as well as the transaction it is currently buffering.
replication_applier_status_by_worker: current status of the thread(s) applying transactions received from the master, provides information about the transactions applied by the applier thread, or by each worker when using a multi-threaded slave.
Using these tables you can monitor information about the last transaction the corresponding thread processed and the transaction that thread is currently processing. This information comprises:
a transaction’s GTID
immediate_commit_timestamp, retrieved from the slave’s relay log
the time a thread started processing a transaction
for the last processed transaction, the time the thread finished processing it
In addition to the Performance Schema tables, the output of
SHOW SLAVE STATUS has three
fields that show:
SQL_Delay: A nonnegative integer indicating the replication delay configured using
CHANGE MASTER TO MASTER_DELAY=N, measured in seconds.
Waiting until MASTER_DELAY seconds after master executed event, this field contains an integer indicating the number of seconds left of the delay. At other times, this field is
Slave_SQL_Running_State: A string indicating the state of the SQL thread (analogous to
Slave_IO_State). The value is identical to the
Statevalue of the SQL thread as displayed by
When the slave SQL thread is waiting for the delay to elapse
before executing an event,
PROCESSLIST displays its
Waiting until MASTER_DELAY seconds after
master executed event.