October 22, 2014

Handling long-running queries in MySQL with Percona XtraBackup

I recently had a case where replication lag on a slave was caused by a backup script. First reaction was to incriminate the additional pressure on the disks, but it turned out to be more subtle: Percona XtraBackup was not able to execute FLUSH TABLES WITH READ LOCK due to a long-running query, and the server ended up being read-only. Let’s see how we can deal with that kind of situation.

In short

Starting with Percona XtraBackup 2.1.4, you can:

  • Configure a timeout after which the backup will be aborted (and the global lock released) with the lock-wait-threshold, lock-wait-query-type and lock-wait-timeout options
  • Or automatically kill all queries that prevent the lock to be granted with the kill-long-queries-timeout and kill-long-query-type settings

Full documentation is here

Diagnosing the problem

Percona XtraBackup has to run FLUSH TABLES WITH READ LOCK before backing up non InnoDB tables. And by default, it will wait forever for this lock to be granted. So you can end up with the following situation:

  • A backup is started while a long-running query is executing
  • The long-running query holds locks that prevent FLUSH TABLES WITH READ LOCK to complete, the backup is stalled
  • All new write queries are blocked, waiting for FLUSH TABLES WITH READ to complete

There are several ways to diagnose the problem. First if you look at the output of innobackupex, you will see that FLUSH TABLES WITH READ LOCK is being requested and that everything is stalled from this point on:

And with SHOW PROCESSLIST, we can see that write queries are waiting for FLUSH TABLES WITH READ LOCK to be released, while FLUSH TABLES WITH READ LOCK is waiting for the long query to finish:

Let’s now use the new options of Percona XtraBackup to have a better control over this global lock.

Aborting the backup if long queries are running

Here is a way to get a long-running query with the employees database:

Now let’s run a backup which will automatically aborts if it cannot take the global lock after waiting for some time:

This command line tells Percona XtraBackup to only wait up to 5 seconds if there is any query running for more than 10 seconds.

If we check the output of innobackupex, we will see something like:

In this case, the backup was aborted because the long-running query did not finish in time. As a side note, because of this bug, it is safer to use Percona XtraBackup 2.1.5+.

Killing long running queries to allow the backup to complete

Let’s change the settings we give to innobackupex. Instead of aborting the backup if it cannot run FLUSH TABLES WITH READ LOCK, let’s now allow the tool to kill any offending query after waiting for maximum 10 seconds:

The output of innobackupex shows that the long-running query is killed and that the whole backup can complete:

Conclusion

The new settings of Percona XtraBackup are nice to avoid nasty interactions between backups and long-running queries. However if you want to use the settings discussed in this article, be prepared for the potential consequences:

  • Allowing the backup tool to kill queries can cause unexpected behavior in the application
  • Aborting one backup can be a minor problem, but if the backup is aborted at each run, this is a major problem!
About Stephane Combaudon

Stéphane joined Percona in July 2012, after working as a MySQL DBA for leading French companies such as Dailymotion and France Telecom.

In real life, he lives in Paris with his wife and their twin daughters. When not in front of a computer or not spending time with his family, he likes playing chess and hiking.

Comments

  1. I find this a recurring and annoying behaviour of FLUSH TABLES WITH READ LOCK.
    However we are so used to issue this query, that we forget we don’t always need it: to get a consistent backup, we only need to flush the MyISAM tables; there is no need to flush any InnoDB table.
    The incentive for using FLUSH TABLES WITH READ LOCK is to get a “point in time” backup: a backup where the server’s binlog position is known; which is useful for setting up new slaves or doing point-in-time recovery. If you don’t need these, you can drop the FLUSH command altogether.

  2. Przemysław Malkowski says:

    Schlomi, this is true, that’s why the –no-lock option exists (http://www.percona.com/doc/percona-xtrabackup/2.1/innobackupex/innobackupex_option_reference.html#cmdoption-innobackupex–no-lock). But flushing is not only needed for non-transactional tables, it is also to avoid inconsistencies when any DDL is executed during the final backup stage (when all the .frm, etc. are being copied).

  3. Fran Garcia says:

    That’s a really welcome addition :)

    What I’m missing from this is some way of limiting the amount of time tables remain locked after the lock has been acquired. For example, say that the backup process acquires the lock, and the copy of all frm files is taking longer than usual (say, because of network issues), so we might want to interrupt the backup and release the lock after 2 minutes to avoid further impact.

    Are there any plans to add a similar feature?

  4. Chandu says:

    Does it only kill the select or if i have a insert or update with select .. from …. .; with a multiple joins and conditions, will it get killed ? would it be a good if its killed or had to wait for them to be rolled back before the extrabackup can proceed further.

  5. Stephane Combaudon says:

    Fran,

    Feel free to open a feature request: https://bugs.launchpad.net/percona-xtrabackup/+filebug

    At the moment, there are different workarounds:
    * Using InnoDB only will minimize the locking time
    * If you don’t need the binlog position, you can use the –no-lock option
    * Using a dedicated backup server allows you to minimize the impact of locking on the application
    * Write a script that monitors the time the lock is held and that will kill the backup if it takes too long

  6. Stephane Combaudon says:

    Chandu,

    If you use –kill-long-query-type=all, any kind of long-running query can be killed, even inserts or updates.

Speak Your Mind

*