This document summarizes and compares several solutions for multi-master replication in MySQL databases: Native MySQL replication, MySQL Cluster (NDB), Galera, and Tungsten. Native MySQL replication supports only limited topologies and has asynchronous replication. MySQL Cluster allows synchronous replication across two data centers but is limited to in-memory tables. Galera provides synchronous, row-based replication across multiple masters with automatic conflict resolution. Tungsten allows asynchronous multi-master replication to different database systems and automatic failover.

1. MySQL
Multi-Master Replication
Challenge
Michael Naumov

2. About Wix
1. We provide a platform to build your own and free awesome web site for any device
2. About 40M Registered Users
3. Around 1.3M new registrations every month
4. About 500 employees with 4 offices around the world.

3. Why do we need Multi-Master topology?
The Vision: Multiple, active DBMS servers with exactly the same data
1. Simple high availability model
2. Operate systems over multiple sites
3. Access geographically “close” data

4. BUT!
It is not so easy to do as its seems

5. MySQL consistency in stand alone
install
Transaction #1 Transaction #2
update users update users
set follower=10 set follower=100
where id=101 where id=101
Stop

6. Consistency in a Distributed environment
Transaction #1 Transaction #2
update users update users
set follower=10 set follower=100
where id=101 where id=101

7. Communication latency
Transaction #1 In US Transaction #2 in EU
Async replication =
No application latency
SemiSync/Sync replication =
Application latency

8. CAP Theorem

9. So what are the solutions out there?

10. MySQL Native Replication

11. How Does Native MySQL M/M
Replication Work
Master 1 Master 2
Binlog Binlog

12. • Statement replication = send client SQL
• Row replication = send changed rows
• Use row replication for multi-master
Row vs. Statement Replication

13. Auto-Increment Key Offsets
• Set different keys on each server to avoid primary key collisions
[my.cnf]
server-id=1
auto-increment-offset = 1
auto-increment-increment = 4
...
[my.cnf]
server-id=2
auto-increment-offset = 2
auto-increment-increment = 4
...

14. MySQL Multi-Master Topologies
Master 1
Master 2 Master 3 Master 2
Master 1
Master-master Circular
Replication Replication

15. MySQLNative Replication Summary
• Built in with well-known capabilities
• MySQL replication by default is asynchronous. In addition to the built-in asynchronous
replication, MySQL 5.5 and above supports semi-synchronous replication.
• Very limited topology support
• Very limited conflict avoidance
• Not a good choice for multi-master if there are
writes to more than 1 master

16. MySQL Cluster (NDB)

17. How Does NDB Work?
Access
Layer
Storage
Layer
Mysql Client Mysql Client
NDB1
NDB2NDB4
NDB3

18. NDB Eventual Consistency Algorithm
• NDB has built-in cross-cluster conflict detection based on epochs and primary keys
• Updates to primary always succeed
• Update on secondary may be rolled back if primary has a conflicting update
• MySQL Cluster resends updates from the primary to “cure” conflicts on the
secondary's clusters

19. NDB Supported Topologies
NDB
Cluster 2
Secondary
NDB
Cluster 2
Secondary
NDB
Cluster 3
Secondary
NDB
Cluster 1
Primary
Master-master Circular
Replication Replication
NDB
Cluster 1
Primary

20. MySQLNDB Cluster Summary
• Allows active/active operation on 2 clusters
• Fully synchronous, no action can be returned to client until transactions on all nodes are
really accepted.
• NDB is mainly an in memory database and also if it support table on disk the cost of
them not always make sense.
• Use horizontal partition to equally distribute data cross node, but none of them has the
whole dataset
• Covers failure of individual MySQL nodes inside the cluster by replication factor
• Detects conflicts automatically on rows

21. Galera

22. How Does Galera Work?
Galera Replicator
Master 1 Master 2 Master 3
wsrep API* wsrep API wsrep API
wsrep API (write set replication API), defines the interface between Galera replication and the DBMS

23. Galera approach is Data Centric
• Data does not belong to a node – Node belong to data. Using global transaction Id’s
• Connect to any nodes for writes
• No headache for auto increment. Galera will do it for you
• Galera replicate the full dataset across ALL nodes.
• Galera replicate data synchronously from one node to cluster on the commit, but
apply them on each node by a FIFO queue (multi thread).

24. Galera Summary
• Galera require InnoDB to work.
• Galera data replication overhead, increase with the number of nodes present in the
cluster.
• Galera do not offers any parallelism between the nodes when retrieving the data;
clients rely on the single node they access.
• Synchronous replication
• Percona XtraDB Cluster is based on Galera

25. Tungsten

26. How Does Tungsten Replication Work
Master
Alpha
Master
Bravo
Remote
Bravo
Slave
Tungsten
Replicator Alpha
Tungsten
Replicator Bravo
Local
Alpha
Master
Local
Bravo
Master
Remote
Alpha
Slave

27. Tungsten Failure Handling
• Replication stops and resumes automatically when network link goes down
• Replication stops on replicator or DBMS failure and recovers after operator restart
• Conflicts can break replication. Auto increment keys should be manually configured
on each node
• Have his own filters for data replication

28. Tungsten Supported Topologies
All Masters
Star
Snowflake

29. Tungsten Summary
• Allows active/active operation on N clusters
• Transfer is asynchronous
• Links can be down for days or weeks if required
• It is the application’s responsibility to ensure there are no conflicts
• Tungsten Replicator can replicate data from MySQL to
MongoDB, Oracle, NuoDB, Vertica and others
• Tungsten allows replication from Oracle by using Change Data Capture (CDC).
Destination DBMS can be MySQL or Oracle

30. Solution Comparison
Native MySQL
(5.6)
MySQL NDB Galera Tungsten
InnoDB + - + +
Asynchronous + - + +
Statement based + + - +
Row Based + - + +
Semi-synchronous + - - -
Synchronous - + + -
Global TRX Id + + + +
Time delay replication + - - +

31. Thank you all
@NaumovMichael
http://goo.gl/OqaigN