Research on Long Transaction Support for HBase

Supporting long transaction in HBase has always been a hot issue since NoSQL is widely approved. In this paper, we first divide the existing methods about long transaction of Key-Value databases into two categories: distributed and centralized, and analyze the distributed methods named Percolator and Haeinsa and the centralized methods named HBaseSI and Omid. Then we propose a non-intrusive long transaction approach for HBase, which bases on the idea of pre-conflict checking and timestamp comparison. We design three algorithms for the proposed approach, namely pre-submit, submit-process and snapshot obtain. We also introduce a timeout mechanism based on collision detection. Experiment results indicate that the proposed method has better performance than other long transaction schemes in the system with more-frequent transaction conflict.