The Cactus Approach to Building Configurable Middleware Services

A number of fundamental abstractions and supporting software mechanisms have been developed for simplifying the problems associated with programming highly dependable distributed systems. For example, transactions provide all or nothing execution despite failures, while ordered atomic multicast supports the replicated state machine approach to fault tolerance by ensuring that changes to the state machine are delivered atomically and in a consistent order despite failures. All of these provide a higher level virtual machine on which to build applications by abstracting away details such as the effect of failures, message reordering and losses, and unconstrained concurrent execution. Each is appropriate for different kinds of applications with different requirements. Many of the concrete realizations of these abstractions have been built as middleware services, i.e., software that is logically layered below the application and above the operating system. For each type of abstraction, however, there are often multiple different ways to define the exact semantics that it provides or other attributes, such as the type of failures it is designed to tolerate. This is especially true for multicast, which have been the basis for a large number of group communication services with varying semantics, including Isis [4], Horus [13], Transis [2], Totem [12], Consul [11], xAMP [14], and RTcast [1]. Often, each is different because it is oriented towards a particular type of application or a given execution environment. Our thesis is that the use of highly configurable software can make it easier to construct applications by simplifying the process of building custom middleware services with specific attributes and guarantees. The availability of an easy route to a custom solution is an attractive alternative to the choices faced by current application builders, which typically involve either trying to use existing services with fixed guarantees or constructing an entire service from the ground up. Here, we describe Cactus, a system that realizes this approach, and discuss some of the challenges in providing configurable services with fault tolerance, real time, and security properties. We then briefly discuss a number of different middleware services implemented using Cactus and outline some of the on-going work in developing new services.