A Theory of Vertically Composable Transactional Objects

We introduce a methodology and formal model that captures the essence of vertically composable transactional objects. Vertical composition adds complexity to transactional systems. As such, we aim to unearth a clean semantic model that strikes a balance between anticipating future implementation methodologies yet, nonetheless, offering a formal treatment of effective existing implementations. To this end we adopt a layered approach and show that first-class treatment of computation reversibility leads to a natural form of vertical composition: a given upper layer in a hierarchy can use inverses to roll back its operations and a contention manager can ensure progress by, at any point, applying inverses on behalf of an executing transaction. The model’s expressiveness is evident, for example, from the fact that we do not require that one object layer use the same implementation strategy (e.g. pessimism versus optimism) as another. Our main technical results are the first proofs of contextual refinement (stronger than serializability) and vertical composition for transactional objects. Our underlying semantics gives rise to a novel transactional variant of Herlihy’s Universal Construction. Our model is a generalization of many known TM implementations, including memory transactions, transactional boosting, some nested transactions, etc. Yet the model also anticipates new strategies and leads to a framework for constructing highly-concurrent systems in a modular way. To this end, we describe how it could be used to implement a highlyconcurrent transactional file system out of linearizable base objects.