Transaction failure tolerance without loss of concurrency

Abstract

Two recovery algorithms are described: one to achieve failure tolerance of sequential transactions that read or write object, and the other to achieve failure tolerance of sequential transactions that manipulate typed objects and act on the outside world. Both require the two-phase transaction model and do not interfere with concurrency. It could be argued that the second one induces overheads in writing data to stable storage. However, it is believed that these overheads can be more easily dealt with than a loss of concurrency.<>