Tuesday, May 15, 2012

Achieving Decoupling: Dependency Injection v/s Event driven design

(Update: A more exhaustive sequel post is in progress on this topic so don't be disappointed by lack of details and stay tuned)

Coupling is considered to be a bad sign in a software design, having seen enough of it in practice I agree with this belief. The need for decoupling has resulted in a cottage industry of IoC containers in Java ecosystem, where it still remains popular, so this is not a philosophical topic.

When I was learning object oriented software design as a student, the gist I received was encapsulation and message passing between objects with state. In theory this sounded great, objects with behavior and encapsulated state seemed unbeatable (given that I only knew BASIC and C).  In real life, however, I have found that it is hard to keep objects simple while keeping their dependency fulcrum in control. This may not correlate in real OO programming languages like Smalltalk but I can't speak for that.

Given that I've only worked on a relatively medium to larger enterprise applications (ranging from 10KLOC to 1MLOC+)  this might sound a bit biased: most enterprise applications are not object oriented. This is counter intuitive given that OO approach is considered to excel at modelling large systems by breaking it down in to more manageable pieces, but I digress.

Coming back to the main issue, how can we control coupling?

Dependency Injection
The most popular option so far is to exploit dependency injection to invert the dependencies. The touted benefits of it are: testability, flexibility and abstracted, and hence, switchable implementations.
For example:
interface IHttp{  InputStream open(URL url); }
class FireFox implements IHttp{ 
 InputStream open(HtttpURL url){  doOpen(url).andLeakSomeMemoryAsDesigned(); } 
}

class Chrome implements IHttp{
 InputStream open(HttpURL url){ feedUserInfoTo(googleAI).thenOpenUrl(url).makeItFast(); }
}

class InternetExplorer implements IHttp{
 //Since no one uses it apart from downloading other browser,
 // being efficient, user friendly or useful is just a misplaced goal.
 InputStream open(HttpUrl url){ 
  waitIndefinitelyOrForeverFor(url).attemptToLoadWhenever(url).possiblyReadOrJustGiveUp(); 
 }
}

class MavenHttpClient implements IHttp{
 InputStream open(HttpUrl url) {
   downloadKnownInternetFiveThousandTimesEveryHour().completelyIgnoringWtfHttp()
   .then()
   .inflictDarkSufferingOnUserAndBeReluctantlyOpinianatedAboutHowYouScrewUpBuilds().asAlways(); 
 }
}

class MyHappyEnterpriseApp{
 IHttp http;
 setBrowser(IHttp browser){this.http = http} //Injected dependency
 void myBusinessLogic(){ http.open("http://blog.nirav.name").doSomething() }
}
So by coding to interface and relying on a dependency injection "container" we get all the prescribed  benefits. This is the de-facto standard to achieve decoupling in current enterprise application architecture in Java ecosystem.

Event Driver Design
The second obvious alternative (may be not so obvious to many) is to fall back to good old message passing which can be synchronous or async. Isn't OO approach all about message passing and keeping state encapsulated while reacting to meaningful messages? Turns out, it doesn't jive well enough with statically typed languages such as Java.

Instead of lethally injecting objects with dependencies (abstract or concrete) why not just send a message which will be observed by objects who are really interested in responding? The benefits here are inclusive of the ones offered by dependency injection but are more evolutionary.

For example, The similar implementation of the above design can be represented as events:
interface IEvent{}
interface HTTPEvents{
 interface OpenUrl extends IEvent{
   InputStream open(HtttpURL url);
 }
}
class FireFox implements HTTPEvents.OpenUrl{
 InputStream open(HtttpURL url){  doOpen(url).andLeakSomeMemoryAsDesigned(); } 
}

class Eventing{
  Dictionary events; //initialize implementations
  T  sync( evt){
    return events.lookupImpl(evt.getType()).send(evt);
  }
}
class MyHappyEnterpriseApp{
 void myBusinessLogic(){ doSomething(Eventing.sync(HTTPEvents.OpenUrl, "http://blog.nirav.name")); }
}

While simplistic, this implementation should provide general idea that dependency injection is not the only true way to achieve decoupling.

Even the most complex systems are designed around the idea of message passing (Linux kernel, Win API and so on) where it successfully achieves decoupling from user land API (POSIX) from physical implementations (e.g. CPU interrupt faults and traps). This elucidates a strong merit that message passing can be applied to a general purpose applications (which are comparatively less complex). It is not a novel approach to system design but it appears that reinventing the wheel is in the fashion as far as enterprise application development goes. JMS has been a solid spec geared more or less towards similar principles but it is viewed only as a queueing mechanism to offload the async task.

I guess not every one sees coupling from the same angle and starts to build yet another money making framework to achieve the same goal. I hope that agent driven libraries like Akka and Erlang style message passing gets some more attention from enterprise app community to see how benefits beyond coupling (performance, fault tolerance and scalability) are achieved with message passing.