C[omp]ute

From: "andrew cooke" <andrew@...>

Date: Fri, 13 Jan 2006 15:53:20 -0300 (CLST)

These are some notes I wrote at work to clarify how messaging with Mule
could (should?) work in our system.  One of the side issues here is that
it's unclear what J2EE buys us.

[...]

Consider two services implemented as:

class edu.noao.nsa.msg.SomeMessage {
  // serializable data transfer object with getters/setters
}

interface edu.noao.nsa.serviceA.ServiceA {
  public SomeResponse doSomething(SomeMessage message);
}

class edu.noao.nsa.serviceA.impl.ServiceA {
  // implements the functionality as a simple bean
}

@Local interface edu.noao.nsa.serviceA.ejb.ServiceALocal
extends ServiceA {}

@Remote interface edu.noao.nsa.serviceA.ejb.ServiceARemote
extends ServiceA {}

@Stateless class edu.noao.nsa.serviceA.ejb.ServiceA
extends ServiceAImpl {
  // do injection from @Resource here
}

class edu.noao.nsa.serviceB.impl.ServiceB {
  private ServiceA serviceA; // plus getters/setters
  public Object someProcess(...) {
    ...
    response = serviceA.doSomething(message);
    ...
  }
}

note the package structure:

edu
 +- noao
     +- nsa
         +- msg
         |   +- serializable message beans
         +- serviceA
         |   +- ServiceA as interface
         |   +- impl
         |   |   +- ServiceA implemented as bean
         |   +- ejb
         |   |   +- wrappers for EJB3
         |   +- test etc
         +- serviceB
             +- ServiceB as interface
             +- impl
             |   +- ServiceB implemented as bean
             +- ejb
                 +- wrappers for EJB3

Now for testing we can write Java unit tests on the impl classes.  We can
even connect together services (using Spring) to test more than one
service at a time.

Note - testing service together may seem more like integration testing,
but in the VOI we follow this same pattern down inside the VOI.  So the
VOI itself is composed of sub-services (components) that plug together as
required.  These then test as described above, but can still be deployed
separately.

The EJB wrappers do two things:
- provide Local/Remote interfaces
- do the necessary injection via @Resource


So as monolithic Java code, that just works.  Either in J2EE or in some
Spring based container.  What extra work needs to be done to deploy the
services on separate machines?

First we need to implement a proxy that intercepts the message from
serviceB.  We already have a set of MessageSender implementations which
send a message to Mule.  So the code is:

class edu.noao.nsa.serviceA.mule.ServiceAMsgProxy implements ServiceA {
  private MessageSender messageSender;
  public void setMessageSender...;
  public MessageSender getMessageSender...;
  public SomeResponse doSomething(SomeMessage message) {
    return (SomeResponse)getSender().sendSynchronous(message);
  }
}


That's it.


Seriously.  No more Java code.  That will send the message object to Mule,
which can be configured to send the message anywhere, via any messaging
system (JMS / convert it to XML / send it as an email / SOAP /
in-memory/VM etc).

This proxy is injected into ServiceB.  ServiceB sees the same interface as
before and talks to it as before.

At the ServiceA end, no extra code is needed at all.  Instead you either:
- deploy ServiceA within Mule, in which case routing is trivial
  (practically automatic)
- deploy ServiceA in EJB, in which case you configure Mule to call the EJB
  (we extend Mule to call EJBs - this was a strange ommission in Mule, but
extension - adding a new "provider" - is quite simple).

So the advantages are:
- Write and test pure Java code
- Type checking of inter-service contracts via interfaces
- No messaging code appears in the "business logic"
- All routing decisions are isolated in Mule
- Only a single, simple proxy at the sender is required
- No interface at all is required at the receiver
- The final deployment structure is not fixed in the design
- Progressively more complex messaging technologies can be configured -
Mule provides synchronous transport over asynchronous messaging

One disadvantage is organising the build so that the cross-dependencies
work OK.  This reflects the cross-dependencies in the contracts that each
service relies on - it's not an artifact of bad engineering, but a
reflection that services do depend on each other via their interfaces, and
that this should be verified as early as possible.

I suspect the best build solution is to use two passes - once to generate
the interfaces and a second time to compile code that depends on all the
interfaces available.

Andrew