Research Of High Available MySQL Cluster Based On Message Middleware

Motivated by the advances of the development of networks, databases and their specific demands in terms of high throughput, low response times, flexible load-balancing, data consistency and fault-tolerance, research on database cluster of high performance and high availability is necessary. By replicating data across the cluster, transactions can access local copies fast. Eager replica control provides data consistency and fault-tolerance in a straight forward way. Furthermore, via an update everywhere approach, all types of transactions can be submitted at any site without restrictions. Despite these characteristics, eager update everywhere replication is rarely used in practice since existing solutions have severe disadvantages in terms of performance and complexity.The current popular database cluster requires special hardware, they must be built upon database engine, and therefore they are not scalable. To solve the problem, we use share-nothing architecture which could construct high-performance and high available database cluster as a database gateway. Based on the group communication provided by a message middleware named spread, we use "read one write all available" protocol. Then write operations are executed on all nodes in the cluster consequently the data consistency is ensured. By setting appropriate ranking of group communication based on atomic multicast and keeping the quantity of messages small, the efficiency of write operations is improved. Message control assures that messages are sent in a same total order, and concurrency control based on the distributed strict two-phase locking protocol ensures the same order of write operations among all sites. In addition, database recovery based on log improves the system availability. Write operations records are added in update log when system is running. During recovery, it acquires shared-mode locks on all relations. Thus, read operations can continue and write sets are delayed. Then, using the log records, entire replication recovery or incremental recovery can make the database on newly joined nodes keep the same to the rest of the cluster.The test result shows that the system acts an eager update everywhere replication tool that provides most of the functionality needed in database cluster. Via an appropriate message ordering, when new nodes join in cluster system, write operation delay is not increased sharply. Read operation delay is nearly twice as long as before, but the system throughput grows linearly when additional nodes added. The scalability is optimistic.