Page History

...

So if a class is updated such that its API is no longer backwards-compatible, it's not possible to use both in the same classpath. The only solution is to change the class name, e.g. by changing its package name - or one could change just the classname.
 \[In both of these cases the code could be made compatible again by keeping the original code alongside the new. However, at some point the old code will probably need to be deleted, at which time the versions will be incompatible anyway.\]

...

If there are multiple versions of the same class on the classpath, there's no guarantee which version will be loaded, so in general the classpath should only ever contain a single version of each class.

What is jar-hell ?

Suppose project A depends on the jars B and C, and these in turn depend on jar D. So the classpath consists of (A,B,C,D). No problem so far.

D is updated to add some new features (call it D2), and jar B is updated to use them (B2). Provided that D2 is binary compatible with D, the a classpath which contains (A,B2,C,D2) will still allow jar C to work correctly.

Now suppose that D2 is not compatible with D, such that jar C does not operate with jar D2. The classpath would have to contain both D and D2. However B2 requires the class from D2, wheres C requires the version of the class from jar D. It's not possible to satisfy both of these. Even for compatible classes in D and D2, there is a problem, because there's no way to determine whether classes will be loaded from D or D2.

Jar-hell is when the classpath needs to contains multiple copies of the same classes in different jars.

Maven dependency resolution

...

This can cause a problem, for which there are several possible solutions - none neither of which are is ideal:

• use relocation POMs
• change the Foo package name
  

...

If the change from commons-foo:commons-foo to org.apache.commons:commons-foo is accompanied by a change to the Java package name, e.g. to org.apache.commons.foo3, then there will be no classpath issue, as both Maven and Java treat the artifacts as different.

There are two possible scenarios here

• The new version of the code is binary incompatible with the old version.
• The new version is binary compatible with the old version.
  

However, the change of Java package name is neither binary nor source-compatible, and can require a lot of work for users of Commons Foo. This may be acceptable if the new version has major incompatible changes to the API, but not otherwise - why should users (who may not even use Maven) be forced to change their code just to upgrade to the latest version (James Carman: the user will thank us when they try to use a library that requires the older version, we shouldn't discount this too much. This approach solves the "jar hell" issue) (Sebb: there is no "jar hell" if the versions are binary compatible)?

For binary-compatible releases, the Java package name should NOT be changed, as that causes unnecessary work for all users. It follows that the Maven groupID should not be changed either, unless relocation POMs are guaranteed to work.

As a concrete example, Logging uses the groupId commons-logging. Changing the package name merely to allow the groupId to be changed would cause an awful lot of work, for almost no benefit.