A few nights ago John Stevenson (@jR0cket) and I were fortunate enough to be taken through a explanation of OSGi by some of its key proponents: Neil Barlett (@njbarlett), Zoe Slattery (@zoe_slattery) and Alasdair Nottingham.
I have certainly battled JAR hell and have also experienced problems with keeping web and enterprise applications isolated from each other (manipulating parent delegation and CLASSPATHs anyone?). This has been throughout my Java career (~10 years now), but I had never investigated OSGi properly as an alternative. Why? Because I never spent the time investigating OSGi properly and I was under the strong impression that it was better suited for IDE, App Server and tooling development as opposed to application development.
So after a good couple of hours of discussion I came out with a greater appreciation for OSGi and the following conclusions.
OSGi's core principles are pretty sound
OSGi is a sound solution for modularity, it clearly breaks the developer away from the artificial JAR model of 'modularity' and instead gets you to think about what Java packages you are dependent on. Through some simple metadata information in your manifest file you can specify the packages that you are dependent on (including a version range). For example, your test code may rely on the org.junit.core package within JUnit. Normally you would reference the entire JAR, but with OSGi you can reference just the package org.junit.core, as well as the version range that you want (e.g. 4.7.1 to 4.8.2). That package can then be supplied by any bundle (which still uses the JAR construct) which 'exports' it. You can probably already envision that a much smaller bundle can be provided than the default large JUnit JAR file.
OSGi actually encourages you to design better APIs
For those following the various clean code or software craftsmanship movements will appreciate the fact that by utilising OSGi, the developer is actually encouraged to think about splitting their public API into loosely coupled 'bundles'. The analogy here is that of Maven, which although polarises developers, does encourage them to use a correct build life cycle (clean, compile, test, package, environmentalise). good API developers should certainly look at supporting OSGi in order to help them designed better APIs
OSGi is relatively easy to get started with
You can start with an existing application and simply provide the OSGi bundle information in the manifest file for that JAR. You will tend to find that a vast majority of the dependencies already have OSGi bundle versions and it is not difficult to swap to an OSGi based build using Maven or Ant. In fact, as OSGi is already a well established technology it probably pays to have an OSGi bundled version of any project artifacts you produce. The tooling is adequate on most major IDEs (especially Eclipse) although the OSGi community would still like to see improvements in this area.
What the OSGi alliance is working on improving
A common/std tool for deploying OSGi applications to runtime environments
OSGi does have a std API for deployments, but this does mean that developers need to implement a deployment mechanism for deploying their apps to runtime environments. Most modern application servers do have OSGi container support, but the deployment mechanism is arguably different each time. This is an area that the OSGi alliance is working on and hopefully some standard tooling around this standard will come out soon. Although it seems like a small usability gap for OSGi, my guess is that this is a real adoption blocker.
Like Java apps in general, OSGi bundles are not Linux/UNIX distro friendly
A major complaint from Linux and UNIX distros over the years has been the fact that Java is not installed according to their packaging standards. Java is also an relatively large installation, which is the source for further criticism. Whilst OSGi certainly helps with the size aspect of Java-based programs, as it is also a late-binding runtime container it does not lend itself to the static, early binding packaging systems that the distros use.
As John pointed out, the existing Iced Tea (OenJDK) project does already produce bundle data that looks a good deal like OSGi bundle data:
Looking into the meta-data of the openjdk-6-jdk package, you can see there is a lot of scope for using OSGi style information.
Maintainer: Ubuntu Developers <email@example.com>
Original-Maintainer: OpenJDK Team <firstname.lastname@example.org>
Provides: java-compiler, java-sdk, java2-sdk, java5-sdk, java6-sdk
Depends: openjdk-6-jre (>= 6b20-1.9-0ubuntu1), libc6 (>= 2.4), libx11-6, zlib1g (>= 1:1.1.4)
Suggests: openjdk-6-demo, openjdk-6-source, visualvm
Conflicts: openjdk-6-jre (<< 6b17~pre3-1), openjdk-6-jre-headless (<< 6b17~pre3-1)
Description: OpenJDK Development Kit (JDK)
However, this does need to be investigated further and perhaps if OSGi and project Jigsaw can interoperate and some tooling is built then I could see bundles being deployed according to distribution standards. This would help Java and OSGi applications gain even more acceptance amongst the Linux/UNIX community. This leas us onto the Jigsaw issue...
The (Potentially) Ugly
Project Jigsaw (Java modularity) is coming as part of JDK 8 and at this stage is not going to work hand-in-hand with OSGi. It is not beyond the realms of possibility that the two will interoperate cleanly in the future and I'm hopeful that the two communities can continue to find common areas despite their fundamentally different approaches.
However, if they do not come together in some simple to use form then it is likely that the Java modularisation space will be fragmented which will only harm the end developer, end user and the Java ecosystem.
You can read more on this topic on a blog post by my co-author for "The Well-Grounded Java Developer" Ben Evans. This post has also been cross posted to the LJC blog.
So are we there yet?
For mainstream adoption by application developers, probably not _yet_. But I think when the tooling gap is closed and as more application developers release OSGi apps, it might just become a de-facto standard for modularity in Java apps. However, there still remains the looming specter of Jigsaw integration/interoperability. There have been signs of the two sides discussing how they can meet in the middle, lets hope that continues.