Summary
At a face-to-face meeting in April 2009, the Fedora committers agreed to begin moving toward OSGi as a module framework for Fedora. Over the remainder of 2009, several of us gained experience with OSGi:
- Eddie worked on a Fedora-on-OSGi project, taking first steps to build the entire Fedora webapp as an OSGi bundle.
- Eddie experimented with building Mulgara as an OSGi bundle.
- Chris changed all Akubra modules to be built as OSGi bundles.
- Bill and Andrew are actively using OSGi in DuraCloud, to aid in the deployment of arbirary services.
When the committers met in London in Feb 2010, we discussed what we have learned about OSGi and came to the conclusion that, while OSGi provides several capabilities that we still see as valuable (dynamic re-configuration, classloader isolation, and a standard module lifecycle), it has a significant learning curve and brings with it a new set of dependency management issues. We also noted that moving to a modern dependency injection framework would not be as difficult, nor necessarily incompatible with an eventual move to OSGi.
So, rather than concentrating on "Moving to OSGi" as a goal in itself, we resolved to improve Fedora's modularity in other, more tractable ways, while using the knowledge we've gained over the last year to increase Fedora's "OSGi-Friendliness" over time. We have identified a few concrete steps we can take to move Fedora's Module Architecture forward:
- Document, maintain, and apply best practices for being "OSGi Friendly"
- Change Fedora's Modules to use a modern dependency injection framework (Spring and/or Guice)
- Retain the long-term goal of having a Fedora OSGi bundle that can be used by other apps
Want to Help?
Great! Please signal your interest by adding your name below, contributing to this wiki page, and participating in related discussions on the fedora-commons-developers mailing list.
Lead: Chris Wilper
Interested Contributors: Dan Davis, Andrew Woods, Asger Askov Blekinge
OSGi Friendliness
What can we do to make Fedora more OSGi-friendly (ideally, without building in any runtime dependencies on OSGi)?
- For jars that are intended to be usable by other projects, build them with the maven-bundle-plugin so they include OSGi metadata.
- Put APIs and implementation details in different packages. Two common conventions in the OSGi world are to use an ".internal" or ".impl" suffix on packages that contain implementation details. If you use the maven-bundle-plugin, such named packages will not be exported by default.
- If possible, avoid using libraries that are known to be problematic in OSGi environments, or require third-party repackaging in order to work in OSGi environments. Well-known examples include:
Dependency Injection Framework
Problem Statement
Fedora's original Server and Module classes were designed in 2002, and provided a common way for major functional components ("modules") of the repository to be plugged in, configured, initialized, and stopped. Problems with the existing framework include:
- It's homegrown. Better, more widely-understood frameworks have come along.
- Unit testing is unnecessarily complicated. The base Module class depends on a Server instance being available in order to function, and the Server base class is not easily mocked.
Requirements
Use standard, well-known frameworks/libraries to:
- Resolve inter-module dependencies via dependency injection
- Provide hooks to initialize/de-initialize modules when the webapp container starts and stops
- Allow re-configuration and plugging in of alternative modules without re-compiling
Non-requirements
This work will NOT attempt to:
- Provide a way to dynamically re-configure modules without restarting
- Provide the ability to run modules in their own classloader space
These capabilities may be added in the future, possibly with the help of OSGi.
Framework Choice
Popular frameworks that support the dependency injection pattern include Spring, PicoContainer, and Guice.
How do they compare? Several articles have been written comparing Spring and Guice, as well as all three. As many have pointed out, Spring and Guice are more than DI frameworks. For our purposes, we should consider the attributes of each that are most relevant to the problem at hand:
|
Spring |
PicoContainer |
Guice |
Supports start/stop lifecycle hooks for components |
Yes (interface or xml-configured) |
Yes (interface or annotation) |
No |
Supports autowiring |
Yes |
Yes |
Yes |
Supports in-code wiring and configuration |
Yes (JavaConfig) |
Yes |
Yes |
Supports external wiring (outside of code) |
Yes (xml) |
No |
Not directly (but it's possible) |
Supports external config (outside of code) |
Yes (xml and/or properties) |
No |
Yes (Names.bindProperties) |
OSGi-Friendly |
Yes (Spring-DM) |
Unknown |
Yes (Guice-Peaberry) |
JSR-330 Support |
Yes, 3.0+ |
In Progress |
In Progress |
Jar Footprint (non-OSGi) |
750kb |
300kb |
650kb |
Implementation Strategy/Principles
- Prefer constructor injection to setter injection
- Minimize coupling of code with DI framework
- Minimize changes to existing Fedora functionality
Implementation Plan
Phase I - Prepare
- List all dependents of existing module interfaces
- Identify and remove unused modules and classes with module dependencies.
- Identify and address circular module dependencies. Refactor if possible; otherwise prepare to use setter injection in these cases.
- Decouple module impls from Module abstract class (as is already done with ResourceIndexImpl/Module and others)
Phase II - Swap
- Decide on DI framework
- Convert fcfg to DI configuration
- Trigger initialization of module singletons via DI framework in context initialization
- Use injected dependencies only, avoiding use of Module and Server at runtime
Phase III - Cleanup
- Get rid of Module, Server, and subclasses
- Get rid of everything else that parses/looks at fcfg