Polygene™
Introduction
Tutorials
Javadoc
Samples
Core
Libraries
Extensions
Tools
Glossary 

What’s an Object anyway?

In OOP the main idea is that we should model our reality by creating Objects. Objects have state, and they have methods. Methods in an object are used to operate on the internal state and understands the domain that is being modeled.

By contrast, in procedural programming the focus is on algorithms, which can use several data structures to perform some task. The focus is on what is going on, rather than the "objects" involved.

With OOP it becomes more difficult to "read" algorithms, as they are spread out in many objects that interact. With procedural programming it becomes difficult to encapsulate and reuse functionality. Both represent extremes, neither of which is "correct". The tendency to create anemic domain models is an indication that we have lost the algorithmic view in OOP, and there is a need for it.

The main flaw of OOP, which COP addresses, is the answer to the fundamental question "What methods should an object have?". In traditional OOP, which really should be called "class oriented programming", the classes tend to have a rather narcissistic point of view. Classes are allowed to dictate what methods are in there - regardless of the algorithms which they are part of - and algorithms then need to be aware of these classes when object instances collaborate in an algorithm. Why? This seems like complete madness to me!! Keep in mind that if there were no algorithms, there would be no need for methods at all!! Algorithms, then, are primary, and objects are what we use as helper structures. In philosophical terms, if there is noone around to observe the universe, there would be no need for the universe itself!

In COP the responsibility for defining the methods is reversed: algorithms which implement interactions between objects get to declare what roles it needs the objects to implement, and the composites can then implement these. For each role there will be an interface, and for each composite wanting to implement a role there will be a mixin in that composite. This mixin can be specific for that composite implementation, or it can be generic and reused. The key point is that it is the OBSERVER of the object, meaning, the algorithm, that gets to decide what the object should be able to do.

This is the same in real life. I don’t get to decide how I communicate with you. I have to use english, and I have to use email, and I have to send it to a specific mailing list. It is the algorithm of the interaction between us, the Polygene™ dev mailing list, that has set these rules, not I as a participant in this interaction. The same should, obviously, be the case for objects.

So, with the understanding that algorithms should define roles for collaborating objects, and objects should then implement these in order to participate in these algorithms, it should be trivial to realize that what has passed for OOP so far, in terms of "class oriented programming", where this role-focus of objects is difficult to achieve, if not even impossible, is just plain wrong. I mean seriously, catastrophically, terminally wrong.

Let that sink in.

The method that has been used so far to get around this has been the composite pattern, where one object has been designated as "coordinator", which then delegates to a number of other objects in order to implement the various roles. This "solution", which is caused by this fundamental flaw in "class oriented programming", is essentially a hack, and causes a number of other problems, such as the "self schizophrenia" problem, whereby there is no way to tell where the object really is. There is no "this" pointer that has any relevant meaning.

The Composite pattern, as implemented in COP and Polygene™, gets around this by simply saying that the composite, as a whole, is an object. Tada, we now have a "this" pointer, and a coherent way to deal with the object graph as though it was a single object. We are now able to get back to the strengths of the procedural approach, which allows the implementer of the algorithm to define the roles needed for the algorithm. The roles can either be specific to an algorithm, or they can be shared between a number of algorithms if there is a generic way for them to be expressed.

Goodness!

The question now becomes: how can we use this insight to structure our composites, so that what is part of the algorithm is not too tightly encoded in the composites, thereby making the algorithms more reusable, and making it less necessary to read composite code when trying to understand algorithms. The assumption here is that we are going to write more algorithms than composites, therefore it has to be easy to ready algorithms, and only when necessary dive down into composite code.

When talking about Composites as Objects in Polygene™ it is most relevant to look at Entities, since these represent physical objects in a model, rather than algorithms or services, or other non-instance-oriented things.

If Entities should implement roles, via mixins, in order to interact with each other through algorithms, then the majority of their behaviour should be put into those role-mixins. These are exposed publically for clients to use. However, the state that is required to implement these roles should not be exposed publically, as they represent implementation details that may change over time, may be different depending on role implementation, and usually has a lot of rules regarding how it may be changed. In short, the state needs to be private to the composite.

This leads us to this typical implementation of an Entity

@Mixins(SomeMixin.class)
interface MyEntity
        extends Some, Other, EntityComposite
{}

where Some and Other are role interfaces defined by one or more algorithms. SomeMixin is the implementation of the Some interface. There is NO interface that is defined by the author of MyEntity. Algorithms first, objects second!

The state needed for these mixins would then be put into separate interfaces, referred to by using the @This injection in the mixins.

interface SomeState
{
    Property<String> someProperty();
}

These interfaces will pretty much ONLY contain state declarations. There might be some methods in there, but I can’t see right now what they would be.

In order to be able to get an overview of all the state being implemented by the Entity we introduce a "superstate" interface:

interface MyState
        extends SomeState, OtherState //, ...
{}

This lets us see the totality of all the state that the Entity has, and can be used in the builder phase:

        EntityBuilder<MyEntity> builder = uow.newEntityBuilder(MyEntity.class);
        MyState state = builder.instanceFor(MyState.class);

        //... init state ...

        MyEntity instance = builder.newInstance();


    }

}

This lets us divide our Entity into two parts: the internal state and the external roles of the domain that the object takes part in. Due to the support for private mixins the state is not unnecessarily exposed, and the mixin support in general allow our role-oriented approach to modeling. The role interfaces are strongly reusable, the mixins are generally reusable, and the state interfaces are usually reusable. This minimizes the need for us to go into the mixin code and read it. If we have read the mixin code once, and the same mixin is reused between objects, then this makes it easier for us to understand it the next time we see it being used in another algorithm.

To summarize thus far, we have looked at why OOP so far has not worked out, why this is the case, and how COP deals with it, and how we can implement a better solution of Entities using Polygene™. All is well!

The next step is to start using these Entities in algorithms. Algorithms are usually stateless, or at least they don’t have any state that survives the execution of the algorithm. There is input, some calculation, and then output. In other words, our notion of services fit perfectly here!

Algorithms, then, should(/could?) be modeled using services. If an algorithm needs other algorithms to compute something, that is, if a service needs another service to do something, we can accomplish this using dependency injection, so that the user of the initial algorithm does not have to know about this implementation detail.

In a "Getting Things Done" domain model, with Projects and Actions, you might then have an algorithm like so for task delegation:

void delegate(TaskExecutor from, Completable completable, TaskExecutor to)
{
   to.inbox().createTask( createDelegatedTask( completable ) );

   completable.complete(); // Delegated task is considered done

   from.inbox().createTask( createWaitingTask( completable ) );
}

In the above I don’t know if "from" and "to" are human users or systems that automatically execute tasks. I also don’t know if Completable is an entire Project or a single Action. From the point of view of the algorithm I don’t need to know! All the algorithm cares about is that the roles it needs are fulfilled somehow. This means that I will be able to extend my domain model later on, and have it be a part of these kinds of algorithms, without having to change my algorithms. And as long as my composites implement the role interfaces, such as TaskExecutor and Completable, they can participate in many different algorithms that use these as a way to interact with the domain objects.

This shows the place of services, as points of contact between objects in a domain model, or more generally, "interactions". These will change often, and will increase in number as the system grows, so it is important that they are easy to read, and that they are easy to participate in. With a focus on roles, rather than classes, this becomes much easier to accomplish!

With the responsibilities of entities, as objects, and services, as algorithms, more clearly defined, the last part to deal with is how these are put together. The services, with the methods now being role-oriented, can obviously be applied to a wide variety of entities, but we now go from general to specific. In our software each general algorithm is typically applied to specific objects in specific use-cases.

How is this done?

This is done by implementing context objects, which pick specific objects and pass them into algorithms. This is typically a UI-centric thing, and as such is difficult to encapsulate into a single method. With the previous example we would need to get the three objects involved, and cast them to the specific roles we are interested in. The "TaskExecutor from" could be the user running the application, the "Completable completable" could be the currently selected item in a list, and "TaskExecutor to" could be a user designated from a popup dialog. These three are then taken by the context and used to execute the "delegate" interaction, which performs all the steps necessary.

The interaction method "delegate" is testable, both with mocks of the input, and with specific instances of the various roles. The implementations are also testable, and if the same mixin is used over and over for the implementation, then only one set of tests is needed for each role interface.

To summarize we have in COP/Polygene™ a way to get the best from procedural and object-oriented programming. As we have seen the functionality falls into three categories, the entities implementing objects, the services implementing interactions, and the user interface implementing the context. This also maps well to the ideas of ModelViewController, which in turn maps well to the new ideas from Mr Reenskaug (inventor of MVC) called DCI: Data-Context-Interaction. As a side-effect of this discussion we have therefore also seen how COP/Polygene™ can be used to implement DCI, which is an important next step in understanding objects and their interactions, the fundamentals of which (I believe) are captured on this page.

That’s it. Well done if you’ve read this far :-)

Comments and thoughts to dev@polygene.apache.org mailing list on this are highly appreciated. This is very very important topics, and crucial to understanding/explaining why COP/Polygene™ is so great! :-)