My webinar “Multi-threaded Replication in MySQL 5.6 and 5.7” on February 25 generated several excellent questions following the presentation (available here for playback along with the slides). I didn’t have time to answer many of the questions during the session and so in this post I answer all of them. Thanks to everyone who attended!
Q: What do you expect from MTS with logical clock? Do you think performance would be good as with per database?
A: MTS with 5.6 is not usable if you have a single database. I do not have numbers, but this is quite frequent. With 5.7 everyone should be able to benefit from multi-threaded replication.
Q: When MySQL 5.6 was released, performance of MTS was lower, than in 5.5, for example. Is this addressed now?
A: I am not sure which specific issue or bug you are referring, but if your data is spread across several databases
Q: How does Percona XtraBackup work with MTS? What are the changes in mysqldump?
A: As long as you are using GTIDs, you can safely take a backup from a slave using multi-threaded replication: with XtraBackup, add the
--slave-info option as usual when taking a backup from a slave and with mysqldump, use
--master-data instead of
Q: For checkpoint position, what if MTS thread apply Insert before creating table where it inserting data. How MTR checkpoint will keep track of these transactions applying by different thread on slave?
A: The worker threads track all execution gaps to make sure that out-of-order execution is safe and to be able to replay all events without forgetting any of them. So it is not possible that a worker thread will insert data in a table that has not been created yet.
Q: Can you use MTS with all binlog_format options?
Q: Is there any way to have the threads work so that no database contention happens?
A: The short answer is no: the goal of the worker threads is to execute the incoming transactions as fast as possible. If that results in database contention, you should probably decrease the number of worker threads.
Q: Why doesn’t multi-threaded replication perform well on a single DB?
A: With 5.6, parallelization is based on isolating the transactions going to each database. If you only have a single DB, no parallelization is possible. You should look at 5.7 and the logical clock algorithm.
Q: Are there any implications with regards to GTIDs and Multi-Threaded replication when running a Master-to-Master setup?
A: I cannot think of any, however I am not sure master-master replication is still very relevant when using GTIDs.
Q: Is there any inconvenience with memory or cache when using more workers than the number of databases?
A: If the number of workers is just a bit higher than the number of databases (like 5 workers for 3 databases), there should not be any issue. However with ridiculously high numbers (500 workers for 2 databases), there might be performance degradation. I have not tested such cases, so I cannot give a good answer there. However the idea is that the number of workers should be close to the number of databases and should exceed the number of cores on the server.
Q: Is there multi-threaded replication in MySQL 5.7?
A: Yes, multi-threaded replication is available in MySQL 5.7 and offers improvements compared to MySQL 5.6 (mainly the parallelization with logical clock).
Q: Have you used DIM_STAT to created load and measure SLAVE Lag? Any interesting take-a-ways from that effort?
A: I used sysbench to generate load and Seconds_Behind_Master from SHOW SLAVE STATUS to measure slave lag. That mainly shows that if your workload is a good fit for MTS (multiple databases and load shared evenly across ), performance benefits can be significant.
Q: Does multi-threaded replication also work with Percona XtraDB Cluster/Percona Server?
A: Percona Server 5.6 is based on MySQL 5.6, so you can use multi-threaded replication exactly as you would use it on MySQL 5.6.
On Percona XtraDB Cluster, it is a bit different: replication inside the cluster uses Galera replication, which has nothing to do with MySQL replication. Note that Galera has offered parallel replication from the beginning (parallel applying of the replicated writesets to be accurate). However if you are using asynchronous replicas, these replicas can benefit from multi-threaded replication if they are running on MySQL/Percona Server 5.6.
Q: What happens to cross db transactions? Do they not replicate?
A: These transactions will replicate, but they will have to wait until all preceding transactions have been executed. Stated differently, cross db transactions introduce serialization, so you should avoid them as much as possible if you want to benefit from parallel applying.
To be accurate, if you have db1, db2 and db3 and if you execute a transaction involving db1 and db2, transactions on db3 can still be applied in parallel. So if you have many databases, cross db transactions may not be that bad.
Q: When using MTS without GTIDs, is “Seconds_Behind_Master” from SHOW SLAVE STATUS valid?
A: Seconds_Behind_Master is based on Exec_Master_Log_Pos. And with MTS, Exec_Master_Log_Pos is not reliable as it indicates the position of the latest checkpoint and not the position of the latest executed transaction. However in practice, checkpoints will happen at least every 300ms by default, so Seconds_Behind_Master is still a good indication of the replication lag. Of course you should keep in mind the usual limitations, such as with multi-tiered replication (if the setup is A->B->C, C will report its lag against B, not against A) or when there is a replication error (then Seconds_Behind_Master is NULL).
Q: How can all the servers be realistically restarted at the same time? There could be a few sec intervals if you have multiple servers [That was when I explained how to enable GTID replication].
A: With MySQL 5.6, the requirements are pretty strict when it comes to enabling GTIDs: all servers must be restarted at the same point in time in the replication stream. As you mention, it is difficult if you have several servers, so the only viable solution is: stop the writes on the master, wait until replication has caught up on all slaves, stop all servers, change the configuration, restart all servers.
What it means is that there is a time range when all servers are down. This is a showstopper for many people, and that’s why Percona Server 5.6 now includes a patch from Facebook that allows an online migration to GTIDs and that’s why MySQL 5.7.6 also offers this option.
* * *
Thanks for all of the great questions – and I hope to see you next month at OpenStack Live and the Percona Live MySQL Conference and Expo 2015 (April 13-16) – both at the Santa Clara conference center in sunny Silicon Valley. Get more info here.