Hitchhiker's Guide to Software Architecture and Everything Else - by Michael Stal

Wednesday, August 31, 2011

The Whole is more than the Sum of its Parts – Creating Complex Systems from Simple Entities

The Internet represents an ubiquitous infrastructure that enables complex systems such as clouds, SOA-based systems, e-mail, or Web 2.0 applications. Despite of its overall complexity, the basic ingredients of the Internet are rather "simple" and easy to understand. Think of parts like TCP/IP, HTTP(S), HTML, URIs, DNS, IMAP, SMTP, REST, or RSS, to name just a few.

Another example though not software-related is Lego. Children (and adults :-) can build the most sophisticated systems from very simple Lego bricks. And even in biological systems like ant populations or in physical systems - like the whole universe itself - we can make similar observations.

Infrastructures such as the Internet with high inherent complexity consist of such simple constituents. Why are we still capable of providing new kinds of innovative software applications based on technologies that are more than 20 to 40 years old? What do all of these examples have in common?

Well, they basically combine simple building blocks with a set of interaction-, adaptation-, and composition-mechanisms. Moreover, they allow to compose higher level building blocks (from simpler parts) that support their own abstract set of interaction, adaptation, and composition. Think of Layered Architecture! And mind the fact that the building blocks in this context are not only static entities but (might) also include behavioral aspects making them active objects.

Interestingly, the composition or interaction of parts leads to cross-cutting functionality that wouldn't be possible otherwise. Most recently, there have been scientific papers claiming that the human personality is "just" a result of cross-cutting behavior - but that what I'd like to mention as an interesting side-remark.

In general, approaches are summarized under emergent behavior or emergence when we speak about (inter)active parts. In this context, I am considering the term "emergent" as a broader concept also including passive ingredients.

But why should we care? What can we software engineers learn from the principle of emergence, evolution or composition?

It is possible to build very complex systems based on simple building blocks. That is very obvious because eventually all physical systems are composed from simple elements. However, we need to address the challenge, how, why and when to apply such composition techniques to our own software. And we also need to address the issue of quality.

While it is theoretically possible to build every system using very small, atomic parts, this would lead to unacceptable development efforts, bad maintenance and low expressiveness. Think about using a Turing Machine for implementing a word processor. Hence, the gap between the behavior we are going to provide in resulting applications respectively artifacts and the basic building blocks used for their construction shouldn't be too wide.

On the other hand, we’d like to support the creation of a whole family of applications in a given domain. (If you are knowledgeable about Product Line Engineering, this should sound familiar.)

There are three fundamental possibilities to construct such systems:

  • Bottom Up, i.e., creating a base of entities that spans a whole domain. We can view the domain as a “language” with the base acting as a kind of “grammar”.
  • Top-Down: i.e., thinking about what kind of artifacts we’d like to create and trying to figure out a set a set of entities from which these artifacts can be composed via one or more abstraction layers.
  • Hybrid: maybe, the most pragmatic approach which combines Top-Down and Bottom-Up.

To make things even more complex, the “grammar” could be subject to evolution which might change the “grammar” and/or the language.

Very theoretical, so far. Agreed, so let me show a practical show case.

A prominent example is the Leader-Followers pattern. Roughly speaking, an active leader is listening to an event source. Whenever the leader detects an event, it turns into a worker and actively handles the event, but only after it selected one of the followers to be the next leader. As soon as the “new born” worker has completed its event handling activity, it turns into a follower. And then the whole story might repeat again. This pattern works nicely for applications such as a logging server. Its beauty stems from the fact that active agents comprise a self-organizing system, leading to non-deterministic behavior.

Engineers often prefer centralized approaches with one or more central  hubs or mediators. That is, we like to introduce a central control. There must be someone or something in control to achieve a common goal, right?

In fact, this assumption is wrong. In some cases, it is necessary to have a control, but this control does not need to be a central component but can be distributed across several entities. Let me give you a real life example:

Some day, you are visiting a shopping mall. After having a nice time, you are leaving the mall. But you forgot where you left the car. Thus, all members of the family start searching in the parking lot until someone locates the car and notifies everyone else. Normally, at least in some families, this search activity would be rather unorganized. Or each family members would agree to search in a specific area, but no one would prescribe how the search pattern looks like. Nonetheless, all share a common goal and will finally succeed. (Let me just assume, the car hasn’t been stolen :-)

An architecture concept that resembles this situation could be active agents that communicate with each other using a blackboard. Just assume, we are building a naive web crawler. On the blackboard you could write down the link where you’d like to start the web crawling. An agent takes the token (i.e., the link) and starts to search the Web page. On the Web page it finds further links which it writes to the blackboard, each link representing a separate token. Now, all active agents can access the blackboard, grab a token, analyze the corresponding page and write the links they find on the blackboard. But, wait a moment, when does the search end? The problem boils down to the question how we handle duplicates (Web sites the system already parsed). For this purpose, we could introduce a hash table where we store pages the agents have already visited. If an agent reads a token which represents an already visited URL, it just throws the token away and tries to grab another one. As soon as there are no more tokens available, the crawling is completed.

Sometimes, we also need (single or multiple) controls. For instance, in most Peer-to-Peer networks, the search for a resource starts using an initial set of controls that aggregate information about their environment. But the search or the registration of resources are conducted in a self-organizing and decentralized way. Note, that the introduction of controls or hubs is not necessary for functional reasons, but helps to increase scalability and performance.

A Sensor (or Robot) Network is an example, where we might even need no control at all. Just consider the case, that we distribute sensors across a whole area to measure environmental data. Even if some of these sensors fail, we are still able to gather sufficient data, given a certain amount of sensors is available. But, of course, one could argue, that there must be someone central who monitors al the sensors. By the way, the network of weather stations is an excellent show case for this strategy.

To summarize this part, in such scenarios with active components, we can have the full range of centralized control, decentralized control, or no control at all. Anyway, the important thing is that each active entity has a predefined goal and there is communication between these entities either using a Peer-to-Peer communication style or one (or more) information hub(s).

As already mentioned, we could add evolutionary elements by letting the agents adapt to changing environments using strategies like survival of the fittest and uncontrolled modification. Genetic algorithms are the best example for this architectural style. Like in neuronal networks, the main challenge for a software engineer is the fact that such systems mostly reveal their functionality using cross cutting and non-deterministic behavior. This makes it difficult to create such systems in a top-down way. Most of the time, such systems are composed bottom-up leveraging a “configurable” evaluation function. With the evaluation function, systems can measure how good the results are and adapt dynamically until the results reach a required level of quality.

Now, I’d like to move to more “conventional” software architectures. As we know, all software architectures are idiomatic, usually providing a stack of idiomatic layers and composition of idiomatic subsystems  and components. (Whatever, “subsystem” and “component” mean in this context :-) Note, that this also holds for patterns and especially for pattern languages.

By  “idiomatic” I am referring to the point that architectural artifacts offer a kind of API which define syntax and semantics from their consumers’ perspective. By composing these artifacts, higher level functionality can be addressed using lower level functionality. Artifacts at the same level cooperate to achieve composite behavior.  Again, we are introducing entities, composition, and communication. Cross cutting behavior addresses composites that reveal specific functionalities or quality attributes. To this end, even quality attributes could be considered functional: they may imply additional functionality or comprise cross cutting behavior.

With other words, in software engineering we face the challenge of hierarchies and composition  respectively integration of languages.

  • The more complex these languages are, the more complex it is to use them.
  • The more complex the compositions of languages are, the more complex it is to integrate them into a whole.

Thus, we need to balance between complexity of these languages and the complexity of their composition. This balance of forces can be achieved by building an appropriate hierarchy of he languages.

Thus, yet another definition of “Software Architecture” could be:

Software architecture is about the hierarchical integration of artifacts that balances the idiomatic complexity of its constituents and the complexity of their composition. In order to meet the given forces, it addresses strategic design decisions using different viewpoints, prepares mandatory tactical aspects, and abides to common guiding principles.

The secret of good design is appropriate partitioning in core parts which are then composed to build a whole such that it can easily address inherent complexity, without introducing accidental complexity. In a platform or product line, this can be really challenging, because we must enable application engineering to create each member of the product family by using the common parts, while allowing to deal with variability. As we are dealing with hierarchies of artifacts, this turns out to be even more complex. When extending the scope of a product line, there might be implications on all aspects of the reference architecture.

The goal of architecture design should be to design complex functionality by composing simple artifacts in a simple way. The tradeoff is between complexity of artifacts and complexity of composition as well as finding an appropriate hierarchical composition. All approaches such as AOP, model-driven design, Component-Based Development, or SOA try to achieve this nirvana. However, introducing basic artifacts and composition techniques is beneficial not sufficient. Instead, it is crucial to cover the problem domain and its mapping to the solution domain.

This is what all the well-known definitions of Software Architecture typically forget to mention :-) 


  • Michael, thanks for this! I agree that architecture is about the forum if the thing being made, or perhaps more accurately, of the purpose it serves; it is not about the tools you use to build it. For example, you don't make a .Net MVC App in terms of the pieces of "an MVC App", you make a flight-booking app in terms of the pieces of booking a flight.

    Christopher Alexander's work addresses the same kind of issues in the built world, and promoted the idea that architecture is about the activities of people and the events that go on in that place, and what goes into making it a good place for such, rather than just making a good "building".

    The software world could learn a lot from his work. In fact, the "patterns" movement was an attempt to do so, but quickly degenerated into patterns of building code rather than of the human functionality that gives software it's usefulness.

    Please have a look at this great video of C.A. addressing the software community on this topic at OOPSLA '96: http://youtu.be/98LdFA-_zfA

    Anyway, I really like your emphasis on the problem domain as well as the solution domain, and see if as a rare example in the same line of thought that C.A. advocates. That is, architecture is about form, and each problem or solution has a form of its own, which cannot be properly captured just in terms of generic software patterns.

    C.A. (in the video above) does, however, outline some of his findings for recurring patterns that always appear in "living structures", e.g. things that can evolve naturally and are an organic fit their domain/purpose, versus gluing together modules cut from a mechanistic process that pays no mind to the domain at hand.

    By Blogger Unknown, at 4:45 PM  

Post a Comment

<< Home