Q: How do I configure a slave if the master is already running and I do not want to stop it?
A: There are several options. If you have taken a backup of the master at some point and recorded the binary log name and offset (from the output of SHOW MASTER STATUS ) corresponding to the snapshot, use the following procedure:
Make sure that the slave is assigned a unique server ID.
Execute the following statement on the slave, filling in appropriate values for each option:
mysql> CHANGE MASTER TO
-> MASTER_HOST='master_host_name',
-> MASTER_USER='master_user_name',
-> MASTER_PASSWORD='master_pass',
-> MASTER_LOG_FILE='recorded_log_file_name',
-> MASTER_LOG_POS=recorded_log_position;
Execute START SLAVE on the slave.
If you do not have a backup of the master server already, here is a quick procedure for creating one. All steps should be performed on the master host.
Issue this statement:
mysql> FLUSH TABLES WITH READ LOCK;
With the lock still in place, execute this command (or a variation of it):
shell> tar zcf /tmp/backup.tar.gz /var/lib/mysql
Issue this statement and make sure to record the output, which you will need later:
mysql> SHOW MASTER STATUS;
Release the lock:
mysql> UNLOCK TABLES;
An alternative is to make an SQL dump of the master instead of a binary copy as in the preceding procedure. To do this, you can use mysqldump --master-data on your master and later load the SQL dump into your slave. However, this is slower than making a binary copy.
No matter which of the two methods you use, afterward follow the instructions for the case when you have a snapshot and have recorded the log name and offset. You can use the same snapshot to set up several slaves. Once you have the snapshot of the master, you can wait to set up a slave as long as the binary logs of the master are left intact. The two practical limitations on the length of time you can wait are the amount of disk space available to retain binary logs on the master and the length of time it will take the slave to catch up.
You can also use LOAD DATA FROM MASTER. This is a convenient statement that transfers a snapshot to the slave and adjusts the log name and offset all at once. In the future, LOAD DATA FROM MASTER will be the recommended way to set up a slave. Be warned, however, that it works only for MyISAM tables and it may hold a read lock for a long time. It is not yet implemented as efficiently as we would like. If you have large tables, the preferred method at this time is still to make a binary snapshot on the master server after executing FLUSH TABLES WITH READ LOCK.
Q: Does the slave need to be connected to the master all the time?
A: No, it does not. The slave can go down or stay disconnected for hours or even days, then reconnect and catch up on the updates. For example, you can set up a master/slave relationship over a dial-up link where the link is up only sporadically and for short periods of time. The implication of this is that, at any given time, the slave is not guaranteed to be in sync with the master unless you take some special measures. In the future, we will have the option to block the master until at least one slave is in sync.
Q: How do I know how late a slave is compared to the master? In other words, how do I know the date of the last query replicated by the slave?
A: If the slave is 4.1.1 or newer, read the Seconds_Behind_Master column in SHOW SLAVE STATUS. For older versions, the following applies. This is possible only if SHOW PROCESSLIST on the slave shows that the SQL thread is running (or for MySQL 3.23, that the slave thread is running), and that the thread has executed at least one event from the master. See the section called “Replication Implementation Details”.
When the slave SQL thread executes an event read from the master, it modifies its own time to the event timestamp (this is why TIMESTAMP is well replicated). In the Time column in the output of SHOW PROCESSLIST, the number of seconds displayed for the slave SQL thread is the number of seconds between the timestamp of the last replicated event and the real time of the slave machine. You can use this to determine the date of the last replicated event. Note that if your slave has been disconnected from the master for one hour, and then reconnects, you may immediately see Time values like 3600 for the slave SQL thread in SHOW PROCESSLIST. This would be because the slave is executing statements that are one hour old.
Q: How do I force the master to block updates until the slave catches up?
A: Use the following procedure:
On the master, execute these statements:
mysql> FLUSH TABLES WITH READ LOCK; mysql> SHOW MASTER STATUS;
Record the log name and the offset from the output of the SHOW statement. These are the replication coordinates.
On the slave, issue the following statement, where the arguments to the MASTER_POS_WAIT() function are the replication coordinate values obtained in the previous step:
mysql> SELECT MASTER_POS_WAIT('log_name', log_offset);
The SELECT statement will block until the slave reaches the specified log file and offset. At that point, the slave will be in sync with the master and the statement will return.
On the master, issue the following statement to allow the master to begin processing updates again:
mysql> UNLOCK TABLES;
Q: What issues should I be aware of when setting up two-way replication?
A: MySQL replication currently does not support any locking protocol between master and slave to guarantee the atomicity of a distributed (cross-server) update. In other words, it is possible for client A to make an update to co-master 1, and in the meantime, before it propagates to co-master 2, client B could make an update to co-master 2 that will make the update of client A work differently than it did on co-master 1. Thus, when the update of client A makes it to co-master 2, it will produce tables that are different than what you have on co-master 1, even after all the updates from co-master 2 have also propagated. This means that you should not co-chain two servers in a two-way replication relationship unless you are sure that your updates can safely happen in any order, or unless you take care of mis-ordered updates somehow in the client code.
You must also realize that two-way replication actually does not improve performance very much (if at all), as far as updates are concerned. Both servers need to do the same number of updates each, as you would have one server do. The only difference is that there will be a little less lock contention, because the updates originating on another server will be serialized in one slave thread. Even this benefit might be offset by network delays.
Q: How can I use replication to improve performance of my system?
A: You should set up one server as the master and direct all writes to it. Then configure as many slaves as you have the budget and rackspace for, and distribute the reads among the master and the slaves. You can also start the slaves with the --skip-innodb, --skip-bdb, --low-priority-updates, and --delay-key-write=ALL options to get speed improvements on the slave end. In this case, the slave will use non-transactional MyISAM tables instead of InnoDB and BDB tables to get more speed.
Q: What should I do to prepare client code in my own applications to use performance-enhancing replication?
A: If the part of your code that is responsible for database access has been properly abstracted/modularized, converting it to run with a replicated setup should be very smooth and easy. Just change the implementation of your database access to send all writes to the master, and to send reads to either the master or a slave. If your code does not have this level of abstraction, setting up a replicated system will give you the opportunity and motivation to it clean up. You should start by creating a wrapper library or module with the following functions:
safe_writer_connect()
safe_reader_connect()
safe_reader_statement()
safe_writer_statement()
safe_ in each function name means that the function will take care of handling all the error conditions. You can use different names for the functions. The important thing is to have a unified interface for connecting for reads, connecting for writes, doing a read, and doing a write.
You should then convert your client code to use the wrapper library. This may be a painful and scary process at first, but it will pay off in the long run. All applications that use the approach just described will be able to take advantage of a master/slave configuration, even one involving multiple slaves. The code will be a lot easier to maintain, and adding troubleshooting options will be trivial. You will just need to modify one or two functions; for example, to log how long each statement took, or which statement among your many thousands gave you an error.
If you have written a lot of code already, you may want to automate the conversion task by using the replace utility that comes with standard MySQL distributions, or just write your own conversion script. Ideally, your code already uses consistent programming style conventions. If not, then you are probably better off rewriting it anyway, or at least going through and manually regularizing it to use a consistent style.
Q: When and how much can MySQL replication improve the performance of my system?
A: MySQL replication is most beneficial for a system with frequent reads and infrequent writes. In theory, by using a single-master/multiple-slave setup, you can scale the system by adding more slaves until you either run out of network bandwidth, or your update load grows to the point that the master cannot handle it.
In order to determine how many slaves you can get before the added benefits begin to level out, and how much you can improve performance of your site, you need to know your query patterns, and to determine empirically by benchmarking the relationship between the throughput for reads (reads per second, or max_reads) and for writes (max_writes) on a typical master and a typical slave. The example here shows a rather simplified calculation of what you can get with replication for a hypothetical system.
Let's say that system load consists of 10% writes and 90% reads, and we have determined by benchmarking that max_reads is 1200 − 2 * max_writes. In other words, the system can do 1,200 reads per second with no writes, the average write is twice as slow as the average read, and the relationship is linear. Let us suppose that the master and each slave have the same capacity, and that we have one master and N slaves. Then we have for each server (master or slave):
reads = 1200 - 2 * writes
reads = 9 * writes / (N + 1) (reads are split, but writes go to all servers)
9 * writes / (N + 1) + 2 * writes = 1200
writes = 1200 / (2 + 9/(N+1))
The last equation indicates that the maximum number of writes for N slaves, given a maximum possible read rate of 1,200 per minute and a ratio of nine reads per write.
This analysis yields the following conclusions:
If N = 0 (which means we have no replication), our system can handle about 1200/11 = 109 writes per second.
If N = 1, we get up to 184 writes per second.
If N = 8, we get up to 400 writes per second.
If N = 17, we get up to 480 writes per second.
Eventually, as N approaches infinity (and our budget negative infinity), we can get very close to 600 writes per second, increasing system throughput about 5.5 times. However, with only eight servers, we increased it almost four times already.
Note that these computations assume infinite network bandwidth and neglect several other factors that could turn out to be significant on your system. In many cases, you may not be able to perform a computation similar to the just shown that will accurately predict what will happen on your system if you add N replication slaves. However, answering the following questions should help you decide whether and how much replication will improve the performance of your system:
What is the read/write ratio on your system?
How much more write load can one server handle if you reduce the reads?
For how many slaves do you have bandwidth available on your network?
Q: How can I use replication to provide redundancy/high availability?
A: With the currently available features, you would have to set up a master and a slave (or several slaves), and write a script that will monitor the master to see whether it is up. Then instruct your applications and the slaves to change master in case of failure. Some suggestions:
To tell a slave to change its master, use the CHANGE MASTER TO statement.
A good way to keep your applications informed as to the location of the master is by having a dynamic DNS entry for the master. With bind you can use nsupdate to dynamically update your DNS.
You should run your slaves with the --log-bin option and without --log-slave-updates. This way the slave will be ready to become a master as soon as you issue STOP SLAVE; RESET MASTER, and CHANGE MASTER TO on the other slaves. For example, assume that you have the following setup:
WC
\
v
WC----> M
/ | \
/ | \
v v v
S1 S2 S3
M means the master, S the slaves, WC the clients that issue database writes and reads; clients that issue only database reads are not represented, because they need not switch. S1, S2, and S3 are slaves running with --log-bin and without --log-slave-updates. Because updates received by a slave from the master are not logged in the binary log unless --log-slave-updates is specified, the binary log on each slave is empty. If for some reason M becomes unavailable, you can pick one slave to become the new master. For example, if you pick S1, all WC should be redirected to S1, and S2 and S3 should replicate from S1.
Make sure that all slaves have processed any statements in their relay log. On each slave, issue STOP SLAVE IO_THREAD, then check the output of SHOW PROCESSLIST until you see Has read all relay log. When this is true for all slaves, they can be reconfigured to the new setup. On the slave S1 being promoted to become the master, issue STOP SLAVE and RESET MASTER.
On the other slaves S2 and S3, use STOP SLAVE and CHANGE MASTER TO MASTER_HOST='S1' (where 'S1' represents the real hostname of S1). To CHANGE MASTER, add all information about how to connect to S1 from S2 or S3 (user, password, port). In CHANGE MASTER, there is no need to specify the name of S1's binary log or binary log position to read from: We know it is the first binary log and position 4, which are the defaults for CHANGE MASTER. Finally, use START SLAVE on S2 and S3.
Then instruct all WC to direct their statements to S1. From that point on, all updates statements sent by WC to S1 are written to the binary log of S1, which will contain exactly every update statement sent to S1 since M died.
The result is this configuration:
WC
/
|
WC | M(unavailable)
\ |
\ |
v v
S1<--S2 S3
^ |
+-------+
When M is up again, you just have to issue on it the same CHANGE MASTER as the one issued on S2 and S3, so that M becomes a slave of S1 and picks all the WC writes it has missed while it was down. Now to make M a master again (because it is the most powerful machine, for example), use the preceding procedure as if S1 was unavailable and M was to be the new master. During the procedure, don't forget to run RESET MASTER on M before making S1, S2, and S3 slaves of M. Otherwise, they may pick up old WC writes from before the point at which M became unavailable.
We are currently working on integrating an automatic master election system into MySQL, but until it is ready, you will have to create your own monitoring tools.