Polygene™ and state modeling

(From Rickard Oberg’s blog, http://www.jroller.com/rickard/entry/qi4j_and_state_modeling, 2009-02-19)

In the Polygene™ project we strive to being able to express our domain as directly as possible, with as little translation as possible between how we think about a domain and how it is actually coded. This then affects our entire view of how application frameworks should be constructed, and what we consider good and bad ways of implementing certain features.

One part which is a good example of this is persistent state modeling. In other approaches and frameworks one would typically use POJOs for the objects, and then plain class fields for references, collections and properties. But during modeling these are not the words we use. If POJOs are used for both Entities and Values, which have radically different semantics, we have to always translate in our heads when talking about them, always keeping mind what the POJO is doing in any particular context. In Domain Driven Design terms, POJOs are not in our Ubiquitous Language.

From a DDD perspective we want to talk about Entities and Values, Properties and Associations. If our code does not reflect this, then there is translation going on, and translation inevitably leads to information loss. In Polygene™, where Composites and not Objects, are the basic construct, we have created specialized forms to model these concepts more explicitly. We have EntityComposites and ValueComposites, each with different ways of creating them and managing them. Both EntityComposites and ValueComposites can then have Properties, as first-class objects, but Properties in ValueComposites are always immutable.


Here’s an example of how you could define an EntityComposite:

interface PersonEntity extends HasIdentity
    Property<String> givenName();
    Property<String> surName();

With these few lines you have done what would have taken a whole lot more effort using POJOs, and what you want to express is very explicit. To define a property you would have had to write a field and two accessors, and if you use interfaces then those accessors would have to be duplicated.

The HasIdentity base interface also includes an identity property for you, as that’s an intrinsic feature of Entities, so that’s all taken care of. So if you speak about Entities in your domain discussions, each having Properties, then you can put that down in code pretty much as-is. This is, to me, a huge advantage over other ways of doing it, including POJO modeling (which lose clarity), UML modeling (which has roundtrip problems), DSL modeling (which lose tools support), and whatnot.


If you want to get picky about it, the above example is probably not how you would model Person. Having a name is just one role that a Person has to play, and since Composite Oriented Programming is all about using roles and context instead of classes you would probably do something like this instead:

interface Nameable
    @UseDefaults Property<String> givenName();
    @UseDefaults @Optional Property<String> surName();

interface PersonEntity extends Nameable, HasIdentity {}

I’ve extracted the ability to be Named to its own interface, and let my domain Entity extend it. This way client code can check for "x instanceof Nameable" rather than "x instanceof PersonEntity", and then do something intelligent with it. By doing this, not only has Nameable become a reusable interface, but the client code that understands it has also become reusable for all domain objects in your model that uses it!

I’ve also marked both properties as @UseDefaults. What does this do? Well, if you have string properties they have the annoying property of being null to begin with, compared to ints and longs which default to 0. We figured that this was such a useful thing that we wanted to be able to mark our properties as being able to have their initial values be set to a default, for the type. For Strings this is the empty string, for primitives they are what you would expect. For Collections they are set to empty collections of the type indicated. "@UseDefaults Property<List<String>> addresses();" would be initialized to an empty ArrayList, for example.

In addition I have set surName() to be @Optional. In Polygene™ we have defined all properties to be not-null as the default, and the same goes for all parameters in method invocations. If you explicitly want to allow properties to be null, so that for example "Madonna" would be an acceptable name, then you can mark it as @Optional. We prefer @Optional to @Nullable since it better expresses the intent from a domain perspective. Avoiding technical terms as much as possible is a another goal (which is damn hard to reach!).


If you want to get really picky about it, not even the above would be a real example. You may want to encapsulate the two properties into one value instead, so that you can more easily perform validation checks when they are updated. What you want are ValueComposites:

interface NameValue
    @UseDefaults Property<String> givenName();
    @UseDefaults @Optional Property<String> surName();

interface Nameable
    Property<NameValue> name();

Normally if you want a property to be immutable, meaning, you can set an initial value but not change it later, you would have to mark the Property as @Immutable. But since NameValue will be assembled as a ValueComposite, which itself is marked with @Immutable, this is implicit for all ValueComposites, and there’s no way to "opt out" of it.

By introducing one more level in the state model you have created an easy way to access the name as a whole and hand it around the system, instead of as two separate properties. Since it is immutable you are also ensured that noone can change it without going through the Entity, and you can also share instances of the name without having to worry about thread-safety.

Privatizing state

The above is already a great step ahead in terms of how you can model your state more easily than having to use POJOs to sort of "fake" the features I’m describing above, and there’s also a ton of cool features and consequences of the whole thing I’m skipping here, for brevity. One of the problems with POJO models that usually come up is that your getters and setters get exposed to clients, and so functionality tend to not be put in the Entities, but rather in services and helper code, thereby scattering the Entity into a bunch of places. What should have been a neat and tidy little package is instead a anorectic little thing whose guts lay splashed around your code, looking all soggy and unappetizing.

What to do about this? One of the great inventions of Polygene™, I believe, is the notion of private mixins. That we can have mixins in an object which are explicitly HIDDEN from usage outside of it. How can we use this for state modeling? What you’d want to do is to model the state of an Entity as a private mixin, which is hidden from clients, and then you write role mixins which map domain methods to that internal state. Here’s an example:

@Mixins( ListablePersonMixin.class )
interface PersonEntity extends Listable, HasIdentity {}

interface PersonState extends Nameable {}

public class ListablePersonMixin implements Listable
    @This PersonState person;

    public String listName()
        String fullName = person.name().get().givenName().get();
        String sn = person.name().get().surName().get();
        if (sn != null) fullName += " "+sn;
        return fullName;

interface Listable
    String listName();

Neat huh? @This is a dependency injection annotation, but rather than the usual generic annotations like "@Inject" and friends, this one actually has a meaningful scope, that is, it requires that "this Entity" implements PersonState. The reference, as it is not extended by PersonEntity, is not visible to clients of the Entity and is hence a private mixin. Furthermore, all of a sudden your client code doesn’t even care if your domain object is Nameable, but rather if it is Listable. Cool! So you can make a UI widget for listing "stuff" that only requires that your thingie is Listable, rather than being a PersonEntity or Nameable.

For extra credit you could move the construction of "fullName" into a method on the NameValue, so that the value is not only a dumb data container, but can also perform useful operations on itself. And for the performance aware, don’t worry, the mixin is lazy-loaded, so if the particular usecase handling the Entity doesn’t need the "Listable" interface the mixin will never be instantiated.

And so much more…

The above is just a small taste of what you can do with state modeling in Polygene™. There is also support for associations and many-associations, the notion of aggregates, and a complete pluggable system for persistence and indexing/querying of data. To end with, here’s a sample of how some other state modeling concepts can be expressed in Polygene™:

interface PersonEntity extends HasIdentity
    Association<PersonEntity> father();
    @Optional Association<PersonEntity> spouse();
    ManyAssociation<PersonEntity> children();
    @Aggregated ManyAssociation<BookNoteEntity> favouriteBooks();

interface BookNoteEntity extends HasIdentity
    Property<String> note();
    Association<BookEntity> book();

I hope they are self-explanatory.

My hope is that with Composite Oriented Programming and Polygene™ we can come one step closer to being able to express our domain as clearly as possible in code.