What's Wrong with the EJB 2 Specification?

The EJB 2.0 beta specification was released with great fanfare last summer during JavaOne. The EJB 2.0 specification introduced new features, including a souped-up version of entity bean CMP, message-driven beans, and additional CORBA interoperability. Application server vendors have rushed to support the EJB 2.0 specification; many of them have quickly provided message-driven bean support, but few have CMP and CORBA interoperability support so far.

Today, the specification is still in public final draft form with no official release date set, which is somewhat uncharacteristic. Previous versions of EJB specifications have typically remained in public final draft form for less than a quarter. So what's going on? Whatever the reason for the stall, I'm glad for it.

I have serious issues with some recommendations made in the EJB specification, most of which revolve around the proposal for introducing a fine-grained "dependent object". Before I ramble on about the nuances of dependent objects, there are other inconsistencies in the EJB specification:

1. There are two tags for primary key identification: <prim-key-class> and <primkey-field>. The former tag is used to indicate the fully qualified class name of the primary key object, and the latter is used to indicate which CMP persistent field is the non-compound primary key of the entity EJB. Why does one tag start with prim-key and the other start with primkey? Also, <prim-key-class> is used for simple and compound primary keys, while the <primkey-field> tag is only used for simple compound primary keys. You must pay close attention to the specification otherwise you might make the silly assumption that use of items in the deployment description would be uniform.

2. The ejbRemove() method is used inconsistently in session and entity beans. For session beans, the ejbRemove() method is invoked when a bean moves from the "method-ready" state to the "does not exist" state. This transition doesn't occur on a predictable basis, even though many developers believe that it occurs every time a client invokes the remove() method on a remote stub. For session beans, the ejbRemove() method is used as a container callback to notify the bean that it's being placed into "inactive" status. It is not used to indicate that the bean has been destroyed, but a container is allowed to make this transition as part of a remove() invocation from the client.

   This contradicts the intended use of ejbRemove() for entity beans, for which ejbRemove() is called when the bean is being destroyed in the persistent store. If you take a close look at the entity EJB state diagram, another container callback is used to notify a bean that it is being moved from the "pooled" state to the "does not exist" state: unsetEntityContext(). As an instructor to hundreds of students who're learning EJBs, it's often a nightmare to explain to them that ejbRemove() for session beans and unsetEntityContext() for entity beans represent the same event transitions, but ejbRemove() for entity beans really means destruction.

   Here's a proposal, though the migration to support it would be difficult: how about changing ejbRemove() in the SessionBean interface to be unsetSessionContext() and alter ejbRemove() in the entity bean interface to be ejbDestroy()? In this scenario, the ambiguous ejbRemove() wouldn't exist in either interface and consistency would exist between setXXXContext() and unsetXXXXContext() callbacks.

Dependent Objects

The EJB 2.0 specification introduced a new approach to doing CMP entity beans. It also introduces the concept of relationships between entity EJBs and dependent objects, which are also new to the specification. In previous versions, all entity beans were considered coarse-grained persistent objects, requiring every access to the bean to be completed over an RMI invocation. The network overhead and the request-level interception work done by a container -- transaction demarcation, security checks, etc. -- made entity EJBs heavyweight objects.

Coarse-grained objects are ideal in many situations, but there are situations in which more finely-grained, persistable objects exist too. For example, a customer profile might be considered a coarse-grained object, but the credit card information stored for a single customer profile might be too fine to warrant having RMI invocations and separate transaction management as part of the object's access logic. A fine-grained object may be needed by a system for one or more of the following conditions:

• the accessed object doesn't need to have its interface exposed to remote clients;
• the accessed object can be accessed over a method invocation as opposed to an RMI invocation (to improve performance);
• the accessed object may not necessarily be identified through a primary key, and duplicates may be possible (they're strictly forbidden with entity EJBs);
• the accessed object follows the life cycle of the accessor object so that it's created and destroyed contemporaneous to its accessor;

Please reference Figure 1: Fine-Grained Objects as EJBs and Figure 2: Advantages of Local Fine-Grained Objects, both below, for another view point.

Figure 1: Fine-Grained Objects as EJBs

Figure 2: Advantages of Local Fine-Grained Objects

The ability for a component architecture to host local, fine-grained objects as "attached children" to remote, coarse-grained objects is critical. If two objects are both considered remote, the persistence manager will ultimately have to perform separate database updates to manage both of the objects. However, if a coarse-grained object is managing a set of finer-grained objects, the chance for a persistence manager to do batch database access can significantly improve performance of the system.

To meet this need, the EJB 2.0 specification defines a dependent object as any object that's created by another object, can only be accessed by another object, or follows the life cycle (creation and destruction) of another object. Dependent objects are not entity EJBs and have a JavaBean-like syntax. Dependent objects can also participate in CMP relationships, giving a variety of options:

• Entity Bean - Dependent Object relationships
• Dependent Object - Dependent Object relationships (the depths of which are unbounded as defined by the specification (!?))
• Dependent objects can participate in one-to-one, one-to-many, and many-to-many relationships
• Dependent objects don't require primary keys
• Dependent objects can have duplicates
• Dependent objects don't have to be "attached" to a parent object. This can occur if a parent object is destroyed and the deployment descriptor is configured not to perform cascading deletions of any child-dependent objects.
• Dependent objects may have cascading deletions (if a parent object that the dependent object is attached to is destroyed, the dependent object and all of its dependent object children will be automatically destroyed)

So What's My Problem?

With all these options available to developers, this must be the right approach to take for incorporating fine-grained objects, right? Wrong. All of this supposed flexibility will ultimately cause more problems than it's meant to solve:

1. Dependent objects are meant to be all things to all people. There are two things that developers currently need to be successful with fine-grained objects in the near future:

   1. Local fine-grained objects that have all of the benefits mentioned earlier -- including local method invocations and bypassing container enhancements, if necessary -- thus allowing developers to model different persistent objects which are to be managed by a single persistence manager employed by a parent, coarse-grained object.

   2. Life cycle objects that follow the same life cycle of a parent object. A life cycle object is created when its parent is created and destroyed when its parent is destroyed. The child object is preferably created and destroyed automatically by the parent's container or persistence manager. This is particularly useful for modeling composition UML relationships where a child object cannot exist without its parent object. In the earlier example, if credit card information is only ever associated with a customer profile, the credit card information may be modeled as a life cycle object where the life of the credit card is managed automatically by the persistence manager of the customer profile object.

2. Dependent objects have a cascading delete deployment descriptor option that instructs a container to manage the destruction of dependent objects attached to parent objects. This is needed to manage life cycle objects. The other characteristic needed to support life cycle objects, an automatic creation mechanism, does not exist, which makes the intentions of the specification developers murky. If the intent is to make dependent objects behave like life cycle objects, which they are according to objectives listed in section 9.3.1 of the EJB 2.0 public final draft specification, then why include cascading delete behavior without including a mechanism for tying the creation of a dependent object to its parent? Presently if you want a dependent object to be created at the same time its parent object is created, you have to create it manually during a creation callback of the parent object. It would have to be done as part of invoking the ejbPostCreate() method of an entity bean that was the parent of a life cycle dependent object.

3. Dependent objects are written differently than entity EJBs. Since dependent objects can only be accessed locally, home and remote interfaces are not necessary. Only an abstract implementation class has to be provided as part of a dependent object's implementation. The container still provides a concrete implementation similar to entity EJBs. The additional syntax that an EJB developer will have to learn will confuse new developers.

I want to emphasize the confusion and ambiguity introduced by dependent objects by raising a final issue. If the point of dependent objects is to standardize local, fine-grained objects, which it must be since dependent objects only partially support life cycle requirements, then they are a small step forward at best. Many developers believe that dependent objects are necessary since there isn't another way to achieve CMP management of a local, fine-grained object.

To the contrary, many application servers have the hooks, flags, and proprietary measures needed to model two entity EJBs in a remote-local fashion. Using these proprietary flags, it's possible to achieve behavior and performance very similar to a dependent object implementation. How? Using BEA's WebLogic Server 6.0 container implementation,

1. implement the parent object and the child object as entity EJBs with CMP;
2. in the WebLogic-specific deployment descriptor set the child EJB to be accessed by reference; (the EJB will be accessed over a local method invocation instead of RMI if the child EJB is in the same virtual machine as its parent);
3. set the transaction attribute of the child EJB to SUPPORTS so that it will inherit any transaction context that the parent EJB uses; (this will insure that any database updates that are performed as a part of the child's ejbStore() callback will be done at the same time that the parent's updates are performed);

Other application servers support similar proprietary flags that achieve the same affect. The disadvantages of this technique are twofold: first, it's proprietary; second, a persistence manager will likely not be able to batch database SQL invocations at the end of a transaction as it would be able to do with a parent object that has a dependent object.

A more detailed analysis of the point of fine-grained objects is obviously warranted. Even without that analysis, the intended use of dependent objects as defined by the specification developers seems fuzzy. Despite addressing a real need identified by developers in the EJB community, this dependent object approach confuses the conceptual issues and ultimately confuses developers too.

Conclusion and What Should Be Done?

1. Get rid of dependent objects.

2. Emphasize the importance of life cycle and local objects. Allow developers to specify objects as life-cycle, local, or both.

3. Develop a technique for fine-grained objects that doesn't introduce more syntax for developers to learn. Consider leaving CMP entity EJBs as the only persistent object implementation, but include deployment descriptor flags that specify fine-grained object.

Is all of this controversial? You bet it is. Despite the number of vendors that have embraced the EJB 2.0 specification, I'm not aware of a single vendor that's implemented dependent objects, including BEA WebLogic and JBoss Server, which are usually the most updated implementations. It's likely that the confusion described here is felt by others in the industry too.

Tyler Jewell , Director, Technical Evangelism, BEA Systems Tyler oversees BEA's technology evangelism efforts that are focused on driving early adoption of strategic BEA technologies into the ISV and developer community.

Return to ONJava.com.

Copyright © 2007 O'Reilly Media, Inc.