Research Of Distributed Transaction In Decentralized Architecture Systems

A decentralized NoSQL system provides highly scalable and available services withno single point of failure or bottleneck of a master node. Yet this kind of system onlyguarantees eventual consistency and the strong consistency should be guaranteed by thedesigner during the application design, which complicates application development. Soit’s necessary to add ACID semantics to NoSQL, which not only retains NoSQL’soutstanding system features, but also help developer avoid maintaining consistency inapplication layer taking advantage of the transaction mechanism. Traditional distributedtransaction’s two phase commit protocol has availability and performance issues. Presentstudies put emphasis on accomplishing the two phase commitment based on Paxos.A configurable decentralized transactional NoSQL called DTCassandra is designedand implemented using Paxos protocol. This system provides the ACID semantics of writetransactions and reads with three consistency levels including current, snapshot andinconsistent read. A configurable consistent read and write algorithm is presented andimplemented which has the following features: First it uses multi-type replicas to separateread and write operation so that different quantities of replicas could be configured;Second it uses log to synchronize write operations across replicas and replicated log fordifferent type of replica is built to implement distributed state machine; Third, it usesPaxos protocol to maintain consensus for witness replicas and NWR quorum protocol forreadable replicas to provide configurable read and write availability with strongconsistency. Making use of data partitioning based on key range and MVCC mechanism ineach partition, fully serializable ACID semantics is provided for local write transactions.2PC has been improved that global transaction across partitions is divided into multiplesub transactions and a primary sub transaction is used as the synchronous point, whichguarantees consistency when node failure happens.4KB data is used as test data for all tests. The results shows that the throughput ofcurrent read is50%higher than Cassandra quorum read while the throughput of localwrite transaction is60%lower than Cassandra quorum write. ACID semantics could beguaranteed by extending Cassandra. High availability is also provided.