Abstract. We propose a novel language construct called communicating transactions, obtained by dropping the isolation requirement from classical transactions, which can be used to model automatic error recovery in distributed systems. We extend CCS with this construct and give a simple semantics for the extended calculus, called TransCCS. We develop a behavioural theory which is sound and complete with respect to the may-testing preorder, and use it to prove interesting laws and reason compositionally about example systems. Finally, we prove that communicating transactions do not increase the observational power of processes; thus CCS equivalences are preserved in the extended language.

We study liveness and safety in the context of CCS extended with communicating transactions, a construct we recently proposed to model automatic error recovery in distributed systems. We show that fair-testing and may-testing capture the right notions of liveness and safety in this setting, and argue that must-testing imposes too strong a requirement in the presence of transactions. We develop a sound and complete theory of fair-testing in terms of CCS-like tree failures and show that, compared to CCS, communicating transactions provide increased distinguishing power to the observer. We also show that weak bisimilarity is a sound, though incomplete, proof technique for both may-and fairtesting. To the best of our knowledge this is the first semantic treatment of liveness in the presence of transactions. We exhibit the usefulness of our theory by proving illuminating liveness laws and simple but nontrivial examples.

A b stra ct. We modify Clean's uniqueness type system in two ways. First, while Clean functions that are partially applied to a unique argu ment are necessarily unique (they cannot lose their uniqueness), we just require that they must be unique when applied. This ultimately makes subtyping redundant. Second, we extend the type system to allow for higher-rank types. To be able to do this, we explicitly associate type con straints (attribute inequalities) with type schemes. Consequently, types in our system are much more precise about constraint propagation. * Supported by the Irish Research Council for Science, Engineering and Technology: funded by the National Development Plan 3 The subscripts of f and g are used only to be able to refer to particular versions of f and g, and are not part of the code.

No abstract

No abstract