Some Thoughts on SOA (and MDSD)
Service-Oriented Architecture, or SOA, is another one of these highly hyped topics in today's IT world. There is no commonly agreed-upon definition of that SOA actually means. Some people equal SOA to "using web services". From our perspective, SOA has nothing to do with certain technologies (WSDL, SOAP, HTTP), but rather constitutes a set of architectural best practices for building large, scalable and composable systems. A well-done component-based architecture with well-defined interfaces and clearly cut component responsibilities can quite justifiable be considered SOA. However, when looking at SOA a bit more closer, one can identify a number of important properties that cannot readily be found in (most) component-based systems:
Service interactions are message-oriented, or document centric. Instead of defining rigidly typed interfaces, document structures (schemas) are defined that serve as the basis for interactions. Done right, this can make evolution of message structures and versioning much simpler.
The interaction patterns, i.e. valid sequences of messages are explicitly defined. Interactions are often conversational, i.e. conversational session state is kept "at both sides" of a service
interaction. These features are the basis for orchestration among services. Usually, interactions are asynchronous, resulting in loosely coupled systems.
Quality of service aspects are explictly addressed. Service providers do not just provide a certain services' functionality, they provide the functionality with a define service level (performance, reliability, etc.).
Service descriptions and characteristics are available at runtime. Using service registries systems can be assembled dynamically.
Often, services are interoperable - they can be used by systems implemented on various platforms.
In addition to these characteristics, services should be designed to be coarse grained and encapsulate functionality relevant from "business perspective" (although nobody can say what this really means). Services are typically (but by no means exclusively) used by explicitly modelled business processes. Finally, like any good IT system, they are secure, transactional and manageable.
Opinions differ whether these characteristics are really so much different from today's well-done enterprise systems. However, what in our view is obvious is that models play a central role in the definition and operation of service-oriented systems.
Message schemas are data structure models that specify required and optional content as well as how message formats change along the evolution of a service.
interaction patterns between services are defined using models; e.g. communicating state machines are useful notations to describe valied sequences of messages as well as exceptional cases.
the levels of quality provided by a service provider (and required by a service consumer) are basically models that are evaluated and checked for compatiblity
The runtime repository basically makes the model information available to runtime queries
And finally, interoperability can be achieved by generating implementations and bindings for various platforms from the same authoritative model.
So the central idea to SOA in our view is: contract first! So the first thing you specify when developing systems is how the communication partners actually interact - and that is independent of ommunication technology and independent implementation platform; rather you specify message formats, interaction patterns and quality of service contracts on an abstract and formal level.
A final word. In the context of SOA you can often see these two pictures where you have this spaghetti-like system with all those interconnections labelled "old" on the one hand side, and a nicely ordered set of components connected through a single bus to which all components are attached labelled "service-oriented" on the other side. This leads people to think that in SOA, all systems must be physically connected through a single technical infrastructure (often web services). Nothing could be more wrong! If you connect all kinds of systems through the same infrastructure you will have a hard time addressing the different non-technical requirements such as throughput, interoperability or performance. The bus you often see on those powerpoint slides must be interpreted as a "logical bus" - i.e. a common, model-based communications infrastructure, where messages can be mapped to various communication technologies and service endpoint be implemented on various platforms