Bundling third party code libraries such as Log4j or Commons Logging is generally easy. Choosing a bundle symbolic name (BSN) for the bundle can be challenging. There are a number of questions that must be asked, options investigated and decisions made. The information here is an attempt to capture some of these issues in an effort to get some level of uniformity within the Eclipse and OSGi communities. This information is put here to generate discussion and surface and resolve issues.
While it is suggested that BSN's be in the form of reverse domain names, this is not required by the OSGi spec. Assuming you want to follow that guideline, there are two choices for the BSN; use your namespace or use the namespace of the library's originator. Some thoughts from the community...
- Using your namespace correctly stamps the bundle as your view of the library
- Using the originator's namespace correctly attributes credit/ownership for the function
- Using the originator's namespace inhibits the originator from free use of their namespace
- Using your namespace leads to many names for the same thing (e.g., org.eclipse.orbit.jetty and org.apache.felix.jetty vs. just org.mortbay.jetty)
As you can see, there are reasonable yet diametrically opposed views. There is some clarity in this space however.
- if you modify or add to the library code then using your own namespace is probably best as you are not incorrectly assigning responsibility (i.e., blame) to the originator.
On the whole the issues raised here only come up because library producers are not producing OSGi bundles. The hope is that over time library producers will change and begin to ship OSGi bundles natively. As such, they will take responsibility for choosing a BSN. Some thoughts from the community...
- When that happens, what if they choose the same name as we choose now?
- The choice we make now will help guide the producers when picking a BSN (we are setting the example)
- What about version number clashes?
- What about people who are using Require-Bundle?
For many libraries it is easy to identify one Java package as being "the root" of the function. This is a reasonable choice for the BSN assuming you are comfortable using the originator's domain name. This option implies names such as org.mortbay.jetty.
Note that some libraries do not use full domain names in their package naming. MX4J packages for example are simply mx4j.*. Following this option would lead to a BSN of just mx4j. As noted above, this is perfectly legal. One could opt here to add in the remaining domain segments and go for net.sourceforge.mx4j.
If you would prefer to use your own namespace (or have modified/added code) a BSN of the form org.eclipse.orbit.jetty seems reasonable. Here the first part is the shortest sequence that is in your control and the remaining part is the "short name" of the originator's function.
To be 100% safe with respect to namespace clashes, you could use the originator's full name in conjunction with your namespace to create a BSN of the form org.eclipse.orbit.org.mortbay.jetty. Here the first part is the shortest sequence that is in your control and the remaining part is the "full name" of the originator's function.
Jeff McAffer says: this seems excessive.
In some cases, there are different implementations of a package or interface, even of the exact same version. In this case, the simple package name and version is not enough to distinguish the different implementations. Note that in Eclipse, it is generally assumed that a bundle symbolic name and version specify a unique set of executable bits, so it would be confusing to distribute bundles with same name, same version, but different implementations. (See bug 226813 for some discussion of this).
javax.mail has been one of the bundles that has several implementors (sun, glassfish, geronimo, etc) so looking at the latest versions of those bundles in Orbit would help illustrate some of these issues and problems.
There are two things that should be done when there are different implementations of the same API from different vendors.
First, 'vendor' attribute should be used when exporting packages, so that clients who care, can specify it on their import package statements. For example, from the javax.mail from glassfish, in the manifest.mf file one export line looks like
javax.mail; vendor=glassfish; version="1.4.1",
For contrast, this could be
javax.mail; vendor=geronimo; version="1.4.1",
The second thing to do is to add the 'vendor' name to the end of the main package name (the main package name being that described in "Option 1" above). So, for javax.mail, this would be javax.mail.glassfish or javax.mail.geronimo. This pattern of the putting the vendor name last is a little bit arbitrary, but seems to make the most sense in the big picture of all bundles, by putting the most significant part of the name first. Here, "most significant" means how someone might be looking for a function in an alphabetized list of bundles. So, to be explicit, it's assumed they are looking for "javax.mail" and then would pick which implementation they wanted.
Note: as of this writing, some of the bundles in Orbit, do not perfectly match the advise given here, simply because they were done before this policy was decided.
Also note: It is assumed we would start to use the vendor name in the bundle symbolic identifier only once there was an actual "conflict" in Orbit. Even if we know there is more than one implementation out in the world, there is no need to use it initially, if there is only one implementation in Orbit. There is no harm in doing so ... we are just stating it is not required.