Fault Tolerant Distributed Computing Using Asynchronous Local Checkpointing
Phillip Kuang, John Field, Carlos A. Varela · 2014
The transactor model, an extension to the actor model, specifies an operational semantics to model concurrent systems with globally consistent distributed state. The semantics formalizes tracks dependencies among loosely coupled distributed components to ensure fault tolerance through a two-phase commit protocol and to issue rollbacks in the presence of failures or state inconsistency. In this paper, we introduce the design of a transactor language as an extension of an existing actor language and highlight the capabilities of this programming model. We developed our transactor language using SALSA, an actor language developed as a dialect of Java. We first develop a basic transactor SALSA/Java library, which implements the fundamental semantics of the transactor model following the operational semantics' transition rules. We then illustrate two example programs written using this library. Furthermore, we introduce a state storage abstraction known as the Uniform Storage Locator following the Universal Actor Name and Universal Actor Locator abstractions from SALSA that uses a storage service to maintain checkpointed transactor states. The transactor model guarantees safety but not progress. Therefore, to help develop realistic transactor programs that make progress, we introduce the Consistent Distributed State Protocol and Ping Director that improve upon the Universal Checkpointing Protocol to aid transactor programs in reaching globally consistent distributed states.