- The client’s view
- Fundamentals of the EJB architecture
- Types of EJB
- Distributed and local EJB semantics
- Anatomy of an EJB
- Principle of operation: session and entity EJBs
- Principle of operation: message-driven EJBs
- The EJB container and its proxies
- Overview of the EJB API
- EJB rules, standards and limitations
- Assembly and deployment
3.4 Distributed and local EJB semantics
Clients of EJBs interact only with the container’s proxies, not the EJB implementation itself. This applies whatever a client happens to be, even if it is itself an EJB in the same application. Typically, the client locates the proxy that implements the factory interface and makes a create() or find() call on it. This factory proxy then provides the client with a reference to a proxy that implements the business method interface, which the client then uses for all subsequent method calls.
In EJB 1.1, all calls on these proxies were considered to be network (RMI) calls, even when EJBs were located in the same JVM. EJB 2.0 allows for the provision of both remote and local (same JVM) proxies, as will be explained later. There was thus only one type of factory interface—called the home interface—and one type of business method interface—the remote interface.
In EJB 2.0, an EJB can be developed to support ‘local’ clients as well as distributed clients. A local client is one which is located in the same JVM as the EJB. This is quite a radical change from the all-RMI strategy of previous EJB versions. This section discusses the changes in the Specification, and the implications these will have for developers. Local clients can be provided with a different client view from distributed clients, and therefore another set of factory and business method interfaces is required. However, local clients still only interact with the EJB through proxies.
The inclusion of support for the local client view was not without controversy and has extremely wide-ranging implications for the developer. To understand why, it is helpful to know a bit about the way the Specification developed.
In Version 1.1 of the EJB Specification and its predecessors, EJBs were, by definition, distributable components. The developer could author EJBs without concern for how they would be physically distributed; all method calls were RMI calls. This model had a certain elegance and purity. The RMI strategy demanded the use of call-by-value semantics, which had the effect of decoupling the caller and the target. The ability to pass references to other objects as arguments is powerful, but makes it more difficult to maintain the caller and the target separately. With strict call-by-value, the author of an EJB could always assume that a method could act upon its arguments without side-effects on the caller.
However, there were problems with this strategy if efficiency was required more than purity. If all method calls were treated as RMI operations, each would involve a network channel set-up operation (e.g., TCP/IP connect), and marshalling and unmarshalling of arguments. Although EJB containers can, and do, act to optimize where calls are known to be within the same JVM, the call-by-value method— required to support RMI operation—prevents efficient argument passing within the same JVM. To obtain good efficiency while maintaining the purity of the application model, we learned to author EJBs with ‘coarse-grained’ client views. Each EJB provided its clients with a small number of complex methods, rather than a larger number of simpler methods. Because fewer method calls were required, the network overhead was kept manageable.
This use of a coarse-grained API was particularly unsuitable for entity EJBs. As we shall see, it is often the case that entity EJBs have relationsips (associations) with other entity EJBs, and these relationships may involve one EJB being dependent on another. As a dependent EJB can only be manipulated through its ‘owner,’ the close coupling between the owner and the dependent requires an efficient method-calling strategy and the use of a fine-grained API.
When EJB 2.0 was in draft, it was proposed that a lightweight calling mechanism be introduced to support the use of dependent entity EJBs. Various possibilities were examined, but in the end, consensus arose in favour of allowing entity EJBs to provide their clients with both a ‘local’ view and a ‘remote’ view. The remote view would be the same as the existing, RMI-based client view, while the local view would be provided by forcing the location of the client and the EJB into the same JVM. Thus EJBs could make intra-JVM calls on one another and thereby pass arguments by reference. For tightly coupled, associated entity EJBs, this is the right thing to do. There was no proposal to weaken the role of the EJB container, and the container proxies would still be interposed between client and EJB.
Having laid the groundwork for the local client view of an EJB, there was no technical reason why it could not be extended to session EJBs. There was, and remains, far less agreement on the validity of this technique with respect to session EJBs. Even purists of component-based development could accept—with some reluctance—the need for a lightweight calling strategy between EJBs that were, by their very nature, tightly coupled. But session EJBs should rarely be tightly coupled to their clients, and the use of local calling semantics is less acceptable here. Not everyone who was involved in the development of the EJB 2.0 specification was agreeable to this change; I myself opposed it quite vocally. Whatever arguments exist on each side, EJB 2.0 does currently support the use of intra-JVM, call-by-reference on session EJBs.
In EJB 2.0, an EJB can be designed to support either distributed clients or local clients or, theoretically, both. The methods exposed to these clients may be the same or different. In all likelihood the methods will be the same if both distributed and local client views are provided, but it is unlikely that a single EJB will need to support both local and remote clients.
Advantages of distributed client-EJB interaction
The EJB server can direct method calls between different servers to provide load sharing. This allows a larger client load to be supported without any particular developer intervention. The ability to do this is one of the most useful features of the EJB architecture, and one that the developer should not sacrifice lightly.
Clients are, on the whole, not affected by whether they are using the EJB’s distributed view or local view. Similar procedures are used to locate, create, and invoke the EJB. However, there are some quite important exceptions that are thrown from the EJB container in the event of an error. The developer needs to be a bit careful about this.
The call-by-value semantics means that the EJB is decoupled from its clients, and the developer does not have to worry about the side-effects of manipulating the arguments passed by clients. Although we can achieve the same effect with a call-by-reference (e.g., the target object can clone its arguments), it relies on the self-discipline of the developer.
Provided that we can find a way to distribute the internal state of an EJB (difficult, but not impossible), the EJB server can provide fault-tolerant access to EJBs. If one EJB server fails, its load can be picked up by another. This is only possible with distributed calling semantics.
Advantages of local client-EJB interaction
Local interaction is much faster, as the overheads of network interaction are significantly reduced.
Despite the closer coupling implied by call-by-reference, sometimes we need call-by-reference for architectural reasons. For example, a client may wish to pass a reference to an output stream to an EJB and have the EJB stream data back as it becomes available. This is very awkward if there is no call-by-reference mechanism.
As all method calls are local, the developer does not have to be concerned with handling exceptions (usually defined as RemoteException and its sub-classes) that may arise at the transport level. For example, the client does not have to deal with the situation that arises if the target EJB is unobtainable because its server is not responding.
3.4.3 Local and distributed interaction: summary
I tend to believe that we (developers) don’t yet have sufficient practical experience with the local client view to understand its full implications. No doubt the outstanding issues will be worked out as more projects start to make use of EJB 2.0 techniques. At present, my feeling is that EJBs should support local clients only for the purpose for which they were originally developed: improving the efficiency of interaction between dependent EJBs and their ‘owners.’ With container-managed persistence, as we shall see, developers have little choice, as the container will expect to see a local client view of a dependent EJB.
It is important to remember that the local client view is available only to clients running in the same JVM as the EJB container hosting the target EJB. In some application servers, the servlet engine may be colocated with the EJB container, and therefore accessible—in theory—to a local client view. In other products, the EJB server may be in a separate JVM, or even on a separate host. This means that, in practice, the local client view is only going to be appropriate for EJB-EJB interactions. In particular, the use of standalone test clients necessitates the provision of a distributed client view.