October 21, 2013 Leave a comment
The project I am working on at the moment contains approximately 1000 interfaces and 4000 classes. We use Autofac for dependency injection, but not all the project uses dependency injection. Within the solution these types are grouped by folder/namespaces and this helps find related types. The only real organization is the use of one Autofac scope to manage sessions. So the types are grouped in two scopes, the global scope and the session scope.
I was expecting the scope separation between classes to be clearly visible on a diagram that represents the dependencies between types. In fact, producing such a diagram turns out to be much harder than I thought. This is a topic I will discuss further in another post. But I am hoping that a well organized diagram would clearly show a natural grouping of classes based on their dependencies.
My goal here is to somewhat formalize this grouping of classes. For this I am proposing to make use of the concept of scoping which plays an important role in many areas of computer science.
Scopes are so pervasive that we use them pretty much unconsciously. In C# we have for example the following scopes:
- statement block, delineated by curly brackets
- using statement scope
- method scope
- class scope
- namespace scope
Some characteristics and properties of scopes:
- Nested or side-by-side
- Non crossing
- Outer scopes have no access to the content of inner scopes
- Inner scopes might or might not have access to the content of outer scopes
- It is also good practice for inner scopes not to modify the state of outer scopes
- Content of outer scopes might be available to inner scopes
- Data passed in on entry
- Data returned on exit
- Creation on entry
- Disposal on exit
- Restoring state on exit
- Lifetime of object inside scope
Generally speaking, scope is considered to be a source code level concept. Scopes are enforced and resolved at compile time. Some aspects are also important at runtime, e.g. call stack management for method calls, disposal of objects when leaving a using scope. But I should emphasize that scopes are always bound to the source code. So the source code of an inner scope is defined within the code of the outer scope. This characteristic represents a strong constraint and limits the usefulness of lexical scopes.
By contrast, a runtime scope defines the scoping of types declaratively rather than lexically. As such types defined in different namespaces or assemblies can live in the same scope. Note that runtime scopes are limited to classes and interfaces. Moreover I will only discuss runtime scopes in the context of (constructor) dependency injection. I will stick to Autofac as I am most familiar with it and as it already provides some support for scopes.
Every interface must define a scope (for the sake of simplicity lets limit ourselves, for now, to a single scope per interface) using a generic marking interface: interface IScoped<TScope> where TScope : IScope.
Scopes themselves are defined as interfaces that derive from the IScope interface.
Finally, scopes can be either singletons (only one can exist within a given scope at any time) or concurrent. I will consider defining other semantics as required, e.g. one-shot singleton or contiguous singletons. The marking interfaces are ISingleScope and IConcurrentScope.
Naturally scopes themselves are scoped.
With Autofac, one registers an implementation of an interface with a class non-declaratively, e.g. by calling the RegisterType method of the builder instance. This has two disadvantages, firstly as the sharing of the class is defined at the registration site rather than declaratively with the interface or class one has to repeatedly navigate between to the registration site and the interface\class implementations. It also makes the analysis of type dependencies by a tool much harder to perform unless one can use Roslyn. Going for a declarative approach solves these two problems.
Sharing, that is whether an instance is shared or not, or in other words whether an instance is a singleton or not can be defined declaratively using either marking interfaces or attributes. In my first version I use marking interfaces: ISingleton (the easy one) and the more dubious IManifold (please, please do suggest a better name).
Let’s have a look at the semantics of runtime scopes.
- A class or scope in TScope can only be resolved in a TScope or a sub-scope of TScope.
- All injected arguments of all constructors of a class (i.e. all its dependencies) in TScope must be in either TScope or a super-scope of TScope.
Condition (1) means that the content of inner scopes is not available to outer scopes, but the content of outer scopes is visible to inner scopes. This last condition could easily be restricted so as to fully isolate the inner scopes. Condition (2) states that all dependencies must come from the same scope or from an outer scope. Here again we could restrict this to only allow dependencies from the same scope.
New scopes are created by existing scopes. A scope can only create a direct inner scope.
Disposal of scopes
Ever scope can be closed, i.e. disposed of. Thanks to Autofac, every class in the scope will also be disposed of if it implements the IDisposable interface.
Just as functions return one or more values on exit, a scope can return one or more values when it is closed. The current mechanism is to instantiate a class in its outer scope using Autofac, so that this shared instance can then be retrieved by the outer scope. I am not really convinced by this mechanism and it doesn’t work well with concurrent scopes. I am investigating other solutions, e.g. continuation scopes or have the closed scope expose the result.