Extending Transaction Management To Capture More Consistency with Better Performance (article invité)

This paper surveys recent work on extended transaction management. It focusses on transaction–oriented applications in large distributed and heterogeneous information systems where applications often span multiple autonomous databases. In such an environment, the challenge is to reconcile data consistency and acceptable performance to the best extent possible, which may imply trading off consistency versus performance in a controlled manner. One approach is to decompose a long user interaction into several independently executed transactions and to provide an additional control sphere as the ”glue” between the transactions. This additional control sphere is usually referred to as an activity or a transactional workflow. Workflow management serves to specify and enforce certain types of execution dependencies between the transactions of a workflow, and thus captures behavioral consistency properties of an application. A second approach is to execute a long user interaction as a transaction, in order to guarantee data consistency, and to relax the isolation and atomicity properties of such a transaction in a controlled manner, in order to achieve acceptable performance. This approach, which is usually referred to as semantic transaction management, is based on exploiting the semantics of the operations that are invoked by the interaction. The paper discusses these two approaches, covering both correctness and implementation issues. A conclusion of the paper is that workflow management and semantic transaction management address complementary issues, and it is suggested that appropriate building blocks from both directions be merged into an integrated architecture.