Afternoon gents,

I'm currently stress testing a Master-Master replication setup using InnoDB as database engine.

We're using this simple script for testing which we run in Linux CLI from a remote server.


while(true) {
    try {
        $conn = mysql_connect('', 'test', 'test');
        if ($conn) {
            $random = rand(0, 1000);
            $res = mysql_query("INSERT INTO test VALUES(0, 'test',    $random)");
            if ($res) {
                echo "\n inserted " . microtime();
            } else {
                echo "\n not inserted " . microtime();
        } else {
            echo "\n can not connect";
    } catch (Exception $ex) {
        echo "\n can not insert"  . microtime();

echo "ok";

The issues we're facing is that we're trying to shut off one of the hosts using nothing but unplugging the power, a hard power off that is.

We're also using MySQL-MMM for fail-over purposes, but that have nothing to do with the issues we're facing but I'll explain the procedure we're using now.

1) Master-Master working perfectly, server1 having virtual IP and is serving writes and reads

2) We shut off server1 by unplugging the cable, the virtual IP gets moved to server2, everything is working and inserts continue after ~20 seconds downtime.

3) We start server1 again, it goes up and gets back the virtual IP address, inserts continues after 1-2 second downtime.

The problem with this is that we lose all inserts that happened during the downtime of server1, and if I type "STOP SLAVE; START SLAVE;" I get these errors:

[ERROR] Slave I/O: Got fatal error 1236 from master when reading data from binary log: 'Client requested master to start replication from impossible position', Error_code: 1236

And if I check the binary log manually with an offset that corresponds with what the mysqld.log file says:

[root@db1 mysql]# mysqlbinlog --offset=623435 db1-mysql-bin.000001
/*!40019 SET @@session.max_insert_delayed_threads=0*/;
# at 4
#121030 12:55:16 server id 1  end_log_pos 106   Start: binlog v 4,   server v 5.1.61-log created 121030 12:55:16 at startup
# Warning: this binlog is either in use or was not closed properly.
ERROR: Error in Log_event::read_log_event(): 'read error', data_len: 112,     event_type: 2
# End of log file
ROLLBACK /* added by mysqlbinlog */;

I understand that the binary log file aren't getting closed properly but aren't InnoDB supposed to take care of this? Surely a hard power off is not something that's very rare, at least not in my mind. I'm running EXT4 filesystem.

This is so far nothing but a lab setup, in reality we run this in state-of-the-art (not submerged ...) data centers with all necessary precautions.

Any light shed on the matter would be greatly appreciated, thank you.


server_id = 2
replicate-same-server-id = 0
auto-increment-increment = 2
auto-increment-offset = 2
replicate-do-db = test
binlog-ignore-db = mysql




I've now switched filesystem from EXT4 to XFS and it took indeed care of the loss of data, but now I have another problem yet quite small and should be easy to solve.

After I go through the procedure to shut off server1, fail over to server2, start server1, roll back to server1. Everything keeps working brilliantly and server1 picks up exactly where server2 left off, the only problem is that server2 stops syncing server1 so it's the other way around.

If I run STOP SLAVE; START SLAVE; it starts syncing and after a few seconds are identical as server1, but why doesn't it do this automatically?

3 Answers 3


Replication and binary logging happen independently from innodb, which can unfortunately cause problems.

Check out: http://dev.mysql.com/doc/refman/5.5/en/replication-options-binary-log.html#sysvar_sync_binlog

From what you describe I suspect that sync_binlog is set to 0 for your servers. Leaving it at 0 means mysql will rely upon the filesystem to handle flushing to disk. Effectively this means that the binlog data will often be in the filesystem cache. That gets flushed to disk by the kernel at some interval, but in the case of a power failure anything in there is lost.

Setting sync_binlog to 1 will force mysql to flush the binlog event to the filesystem using fdatasync after every commit. This is safer (since you will lose 1 transaction at most in case of power failure), but creates a lot more disk activity. Benchmark and see what the impact is for your workload. Knowing the tradeoffs for both scenarios will hopefully help you make an informed decision.

Hope that helps.

  • Hi Drogart, thank you for your input but it seems even with binlog sync set to 1 I get data loss, is MySQL native replication really broken and can't handle power loss? I must say that's a surprise.
    – Lars
    Commented Oct 31, 2012 at 8:31
  • Is your data loss a single transaction? That is what I would expect if it is flushing the binlog at each commit.
    – drogart
    Commented Oct 31, 2012 at 17:16

After some thought I figured out that the slave can't tell the difference between a silent (idle) master from a crashed one unless an error is given, so the TCP connection will stay online until the slaves default TCP time out is reached.

On my boxes, this was the default

[root@db1 ~]# sysctl -e net.ipv4.tcp_keepalive_time
net.ipv4.tcp_keepalive_time = 7200

So I sorted it out with

echo 30 > /proc/sys/net/ipv4/tcp_keepalive_time
echo 2 > /proc/sys/net/ipv4/tcp_keepalive_probes
echo 1 > /proc/sys/net/ipv4/tcp_keepalive_intvl

This makes it send SYN packages every 30 seconds to see if the connection is still replying, and will bring up the sync way sooner. Don't forget to restart the MySQL process so the sockets gets reopened, and add it to rc scripts so the settings stay after a reboot.

There's a lot of knobs to tweak for this so here's a URL with more information on the subject. http://dom.as/2006/09/12/mysql-tcp-network-tuning/

If you're interested in how to set this up, read here. http://mute.nu/2012/how-to-set-up-a-2-node-apache-and-mysql-fail-over-environment/

So all in all, put the MySQL data directory on XFS and tweak the TCP keep alive settings and you're all set.


Run this, and see what changes might eliminate the need for STOP/START SLAVE:


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.