The-Spirit | WebGL Experiment - Live demo | Video | Graphics library

In line with the Open-Closed Principle, I typically design my Java packages and libraries in such a way that there is a generic "interface" or "API" package/library and one or more implementations (quite similar to many common APIs like JDBC or JAXP/SAX). To locate an implementation (or sometimes multiple implementations) in the base API library without violating OCP, I commonly use Java's ServiceLoader mechanism, or occasionally classpath scanning via third-party libraries like ClassGraph or Reflections. From a Maven perspective, the implementations are brought in as runtime dependencies (as they're only needed at execution time, but not at compile time). Pretty standard stuff.

So, now, I want to make some of these packages available as OSGi bundles (with API and implementation in separate bundles), but since in OSGi each bundle has its own class loader, neither classpath scanning nor the ServiceLoader API will work for this purpose. At first glance, OSGi's "fragment" mechanism seems to be the closest equivalent to the plain-Java setup described above. In that scenario, the API bundle would be the "fragment host", and concrete implementations would attach as fragments to that host bundle. As the fragment host and all its attached fragments use the same class loader, the standard plain-Java mechanisms like ServiceLoader or ClassGraph would conceivably still work. This would also have the advantage that there would be no need to detect whether a library/bundle is running in an OSGi context or not, and no OSGi framework dependencies are needed.

So, in a nutshell, my question is: are fragments the correct way to implement runtime-only dependencies in OSGi or is there a better (or more standard) way? Preferably, I'm looking for a solution that works in an OSGi container but does not require a dependency on OSGi itself.