The ultimate gist of DDD

The bounded context

Tweaking the business language and renaming classes and methods is tricky, but thanks to integrated development environment (IDE) features and plug-ins, it is not terribly problematic. However, failing to identify subdomains that are better treated independently could seriously undermine the stability of the whole solution.

No matter how hard you try, your UL will not be a unique set of definitions that is 100-percent unambiguous within your organization. In fact, the same term (for example, customer) might have different meanings across different business units. Like suitcases on an airport baggage belt that look alike, causing confusion among travelers, functions and names that look alike can cause problems in your solution.

Understanding differences between functions and names is crucial, and effectively addressing those differences in code is vital. Enter bounded contexts.

Making sense of ambiguity

When analyzing a business domain, ambiguity occurs. Sometimes we run into functions that look alike but are not the same. When this occurs, developers often reveal an innate desire to create a unique hierarchy of highly abstracted entities to handle most scenarios and variations in a single place. Indeed, all developers have the secret dream of building a universal code hierarchy that traces back to a root Big-Bang object.

The reality is that abstraction is great—but more so in mathematics than in mere software. The great lesson we learn from DDD is that sometimes code fragmentation (and to some extent even code duplication) is acceptable just for the sake of maintenance.

The cost of abstraction

Abstraction always comes at a cost. Sometimes this cost is worth it; sometimes it is not.

Originally, abstraction came as a manna from heaven to help developers devise large domain models. Developers examined a larger problem and determined that it could be articulated as many smaller problems with quite a few things in common. Then, to combat code duplication, developers righteously added abstraction layers.

As you proceed with your analysis and learn about new features, you might add new pieces to the abstraction to accommodate variations. At some point, though, this may become unmanageable. The bottom line is that there is a blurred line between premature abstraction (which just makes the overall design uselessly complex) and intelligent planning of features ahead of time. In general, a reasonable sign that abstraction may be excessive is if you catch yourself handling switches in the implementation and using the same method to deal with multiple use cases.

So much for abstraction in coding. What about top-level architecture? With this, it’s nearly the same issue. In fact, you might encounter a business domain filled with similar functions and entities. The challenge is understanding when it’s a matter of abstracting the design and when it’s breaking down the domain in smaller parts. If you break it down in parts, you obtain independent but connected (or connectable) functions, each of which remains autonomous and isolated.

Using ambiguity as the borderline

A reasonable sign that you may need to break a business domain into pieces is if you encounter ambiguity regarding a term of the UL. In other words, different stakeholders use the same term to mean different things. To address such a semantic ambiguity, the initial step is to determine whether you really are at the intersection of two distinct contexts. One crucial piece of information is whether one term can be changed to a different one without compromising the coherence of the UL and its adherence to the business language.

An even subtler situation is when the same entity appears to be called with different names by different stakeholders. Usually, it’s not just about having different names of entities (synonyms); it often has to do with different behaviors and different sets of attributes. So, what should you do? Use coding abstractions, or accept the risk of some duplication? (See Figure 2-3.)


FIGURE 2-3 Domain and subdomains versus domain models and bounded contexts.

Discovering ambiguity in terms is a clear sign that two parts of the original domain could possibly be better treated as different subdomains, each of which assigns the term an unambiguous meaning. DDD calls a modeled subdomain a bounded context.

The savings of code duplication

From long experience in the code trenches, my hearty suggestion is that whenever you feel unsure whether abstraction is necessary, then by default, it isn’t. In that case, you should use code duplication instead.

That said, I know that tons of articles and books out there (including probably a few of mine) warn developers of the “don’t repeat yourself” (DRY) principle, which encourages the use of abstraction to reduce code repetitions. Likewise, I’m also well aware that the opposite principle—write every time (WET)—is bluntly dismissed as an anti-pattern.

Yet, I dare say that unless you see an obvious benefit to keeping a piece of the top-level architecture united, if a term has ambiguity within the business language that can’t just be solved by renaming it using a synonym, you’d better go with an additional bounded context.

In coding, the cost of a bad abstraction is commonly much higher than the cost of duplicated code. In architecture, the cost of a tangled monolith can be devastating, in much the same way the cost of excessive fragmentation can be. Yes, as usual, it depends!

Devising bounded contexts

A bounded context is a segment of the original model that turns out to be better modeled and implemented as a separate module. A bounded context is characterized by three aspects:

  • Its own custom UL

  • Its own autonomous implementation and technology stack

  • A public interface to other contexts, if it needs be connected

As a generally observed fact, the resulting set of bounded contexts born from the breakdown of a business domain tends to reflect (or at least resemble) the structure of the owner organization.

Breakdown of a domain

Here’s an example taken from a realistic sport-tech scenario. If you’re called to build an entire IT system to manage the operations of a given sport, you can come up with at least the partitions in subdomains shown in Figure 2-4.


FIGURE 2.4 Breakdown of an example domain model in a sport-tech scenario.

It’s unrealistic to build the system as a single monolith. And it’s not a matter of faith in the software creed of microservices; it’s just that, with a decent analysis of the domain, processes, and requirements, you’ll see quite a few distinct clusters of related operations (although maybe not just the six shown in Figure 2-4). These distinct blocks should be treated as autonomous projects for further analysis, implementation, and deployment.

In summary, each bounded context is implemented independently. And aside from some technical resources it may share with other contexts (for example, a distributed cache, database tables, or bus), it is completely autonomous from both a deployment and coding perspective.

Shared kernels

Suppose you have two development teams working on what has been identified as a bounded context and you have an agreed-upon graph of functionalities in place. At some point, team 1 and team 2 may realize they are unwittingly working on the same small subset of software entities.

Having multiple teams share work on modules poses several synchronization issues. These range from just keeping changes to the codebase in sync to solving (slightly?) conflicting needs. Both teams must achieve coherency with their respective specs—not to mention any future evolutions that might bring the two teams into a fierce contrast. (See Figure 2-5.)


FIGURE 2.5 Discovering a shared kernel.

There are three possible ways to deal with such a situation. The most conservative option is to let each team run its own implementation of the areas that appear common. Another option is to appoint one team the status of owner, giving it the final word on any conflicts. As an alternative, you could just let the teams come to a mutual agreement each time a conflict arises. Finally, there is the shared kernel option.

Shared kernel is a special flavor of bounded context. It results from a further breakdown of an existing bounded context. For example, the subdomain in Figure 2-5 will be partitioned in three contexts—one under the total control of team 1, one under the total control of team 2, and a third one. Who’s in charge of the shared kernel? Again, the decision is up to the architect team, but it can be one of the existing teams or even a new team.

Legacy and external systems

For the most part, bounded contexts isolate a certain related amount of behavior. Identifying these contexts is up to the architect team. However, certain pieces of the overall system should be treated as distinct bounded contexts by default—in particular, wrappers around legacy applications and external subsystems.

Whenever you have such strict dependencies on systems you don’t control (or are not allowed to control), the safest thing you can do is create a wrapper around those known interfaces—whether a plain shared database connection string or an API. These wrappers serve a double purpose. First, they are an isolated part of the final system that simply call remote endpoints by proxy. Second, they can further isolate the general system from future changes on those remote endpoints.

In DDD jargon, the isolated wrappers around an external system are called an anti-corruption layer (ACL). Simply put, an ACL is a thin layer of code that implements a familiar pattern. It offers your calling modules a dedicated and stable (because you own it) programming interface that internally deals with the intricacies of the endpoints. In other words, the ACL is the only section of your code where the nitty-gritty details of the remote endpoints are known. No part of your code is ever exposed to that. As a result, in the event of breaking changes that occur outside your control, you have only one, ideally small, piece of code to check and fix.

Coding options of bounded contexts

How would you code a bounded context? Technically, a bounded context is only a module treated in isolation from others. Often, this also means that a bounded context is deployed autonomously. However, the range of options for coding a bounded context is ample and includes in-process options.

The most common scenario—and the most common reason for wanting a bounded context—is to deploy it as a standalone web service accessible via HTTPS and JSON, optionally with a private or shared database. A bounded context, though, can easily be a class library distributed as a plain DLL or, better yet, as a NuGet package. For example, it is almost always a class library when it represents the proxy to an external system.

The public interface of a bounded context with other bounded contexts can be anything that allows for communication: a REST or gRPC gateway, a SignalR or in-process dependency, a shared database, a message bus, or whatever else.