Monday, August 26, 2013

Dear Prudence: Cant we achieve the modularity through normal jars?

Dear Prudence I've some doubts related to osgi I am new to OSGI framework. I was going through the sites and read about OSGI framework. Frankly speaking I did notunderstand anything. Following are my doubts
  • OSGi is supposed to provide modularity. Cant we achieve the modularity through normal jars?
  • What does it mean that OSGi has a dynamic component model?
  • Bundles can be installed,started,stopped,updated,etc. Why do we want to install the bundles? Why cant we access directly like what we access other normal jars?
I am totally confused. Can somebody answer me ? If it is possible to give some examples also?

Dear Confused,
Your question first puzzled me a bit since there is so much documentation on the Internet today and there are plenty of books that take you from minute detail to broad overview. Not to talk about the hundreds of 'hello world' tutorial blogs. Then it dawned on me that many of these tutorials seem to start with explaining why the author felt compelled to make this tutorial because OSGi was so much easier and more powerful than those other 99 blogs that were read before OSGi was understood ... Maybe there is something in OSGi that makes it really hard to understand before you know it.

I guess everybody has a bubble of knowledge that makes it hard to learn/understand anything outside that bubble. I know first hand, last year I really learned Javascript and found myself balking at seemingly bizarre and complex patterns until they became obvious. Your question seems to indicate that your knowledge bubble does not intersect with the bubbles of people advocating OSGi. So lets design a module system based on normal JARs.

I guess we should start with defining what a module is and why we need it. A software module is characterized by having a public API that provides access to a private implementation. By separating the API from the implementation we can simplify the usage of our module since an API is conceptually smaller than the API + implementation and therefore easier to understand. However, the greatest benefit of modules comes when we have to release a new revision. Since we know that no other module can depend on private implementation code we are free to change the private code at will. Modules restrict changes from rippling through the system, the same way as firewalls restrict fires.

Lets make a framework that can use a JAR as a module. Best practice in our industry is to make code private by default, that is no other module can access the inside.This is the standard Java default, without the public keyword fields, methods, and classes revert to being only locally accessible.

However, if nobody outside the JAR can see anything of the inside then this code can never be called. Like Java, we could look for a class with a public static main(String args[]) method in this module to start the module. Since we do not want to search all classes in a module to find this main class (which also means we could end up finding multiple) we need a way to designate the JAR's main class. Such a mechanism is already defined by Java in the JAR specification: the Main-Class header in the JAR's manifest (a text file with information about the JAR). So we could call such a designated main method in each module to start the module, it can then run its private code. However, the main method does not allow us to stop the module. So let's create a new header for this purpose and call it Module-Activator. The class named in the Module-Activator header must then implement the ModuleActivator interface. This interface has a start and stop method, allowing the framework to start and stop each module.

If the private code executes it will likely need other classes that are not in the JAR. Our framework could search the other modules for this referred class if we knew which part of a module was private and which part was public. Since Java already has a namespace/accessibility/encapsulation hierarchy, we should try stay close to these concepts. This Java hierarchy is field/method, class, package and the package is therefore the best candidate to designate private or public. Since there is currently no Java keyword defined to make a package private or public we could use annotations on a package. However, annotations mean that we need to parse the JAR before we can run it, this is in general a really bad idea for performance and other reasons. Since we already defined a header for the activation, why not define another header: Module-Export? This header would enumerate the packages in the JAR that provide the public API.

This minimalistic module system (only two headers) is very attractive but there is a catch that becomes apparent when you build larger systems.

What happens when you have multiple modules in your framework that both provide a given class but their revision differs? The standard Java approach is to do a linear search of the modules and the first module that declares a class wins. Obviously you should never design a system that has multiple revisions of the same class. However, larger systems tend to run into this problem because dependencies are transitive. JARs depend on other JARs, that depend on further JARs, ad nauseum. If, for example, many of these JARs depend on Log4j it is easy to see that not all these JARs will use the same version of Log4j. For a simple library as Log4j that is generally backward compatible you want to have the latest revision but for other libraries there is no guarantee that revisions are backward compatible.

Basically ignoring this erroneous situation like Java (and Maven) does is not very Java-ish. We can't have a system that fails in mysterious ways long after it was started; in Java we generally like to see our errors as early as possible.  For example, the Java compiler is really good in telling you about name clashes, why should we settle for less on the class path?

Since we already export the packages that are shared with the Module-Export manifest header, why not also specify the imported packages in the manifest with a Module-Import header? If the import and export headers also define a version for the package then the framework can check a-priori if the modules are compatible with each other before any module is started, giving us an early error when things are not compatible. You could do even better. We could also make sure that each module can only see its required version of a package, allowing multiple revisions of the same module in one system (this is generally considered the holy grail against JAR hell).

So dear Confused, we've just designed a simple module system based on plain JARs and three manifest headers to describe the module. Fortunately we do not have to struggle to mature this simple design for lots and lots of very subtle and complex use cases because such a module system already exists: its actually called OSGi.

Let me answer your dynamics questions next week,

Peter Kriens

1 comment:

  1. You might consider adding the last missing piece. When the module initializes its private classes it will have to make them visible to the outside. So it will have to register the implementation class in a central registry most probably by using the public interface and some other properties.