Interactive Transaction Processing for In-Memory Database System

In-memory transaction processing has gained fast development in recent years. Previous works usually assume the one-shot transaction model, where transactions are run as stored procedures. Though many systems have shown impressive throughputs in handling one-shot transactions, it is hard for developers to debug and maintain stored procedures. According to a recent survey, most applications still prefer to operate the database using the JDBC/ODBC interface. Upon realizing this, the work targets on the problem of interactive transaction processing for in-memory database system. Our key contributions are: (1) we address several important design considerations for supporting interaction transaction processing; (2) a coroutine-based execution engine is proposed to handle different kinds of blocking efficiently and improve the CPU usage; (3) a lightweight and latch-free lock manager is designed to schedule transaction conflicts without introducing many overhead; (4) experiments on both the TPC-C and a micro benchmark show that our method achieves better performance than existing solutions.