Abstract
The deterministic database systems have recently been shown to yield high throughput on a cluster of commodity machines while ensuring the strong consistency between replicas, provided that most of the data can be fit in the memory on these machines. However, standard OLTP data is unlikely to fit in the memory of commodity machines. The long transactions induced by the buffer swapping have a critical impact to the throughput of deterministic database systems. In this paper, we present IDA, a redesigned transaction execution architecture which reduces the occurrence and impact of long transactions. IDA redesigns the buffer layer with versioned record-based buffer pool and eliminates the need of concurrency control layer. The storage layer is redesigned to take advantage of sequential writes and simplify the recovery layer. We implement IDA on a real database system. Experiments are conducted and the results demonstrate the effectiveness of IDA.