I work in a small shop where we have a LOT of legacy Cobol code and where a methology has been adopted to allow us to minimize forking and branching as much as possible.

For a given release we have three levels:

• CORE - bottom layer, this code is common to all releases
• GROUP - optional code common to several customers.
• CUSTOMER - optional code specific for a single customer.

When a program is needed, it is first searched for in CUSTOMER, then in GROUP and finally in CORE. A given application for us invokes many programs which all are looked for in this sequence (think exe files and PATH under Windows).

We also have Java programs interacting with this legacy code, and as the core-group-customer lookup mehchanism does not lend it self easily to Java it has tended to grow in a CVS branch for each customer, requiring much too much maintainance. The Java part and the backend part tend to be developed in parallel.

I have been assigned to figure out a way to make the two worlds meet.

Essentially we want a Java enviornment which allows us to have a single code base with sources for each release, where we easily can select a group and a customer and work with the application as it goes for that customer, and then easily switch to another codeset and THAT customer.

I was thinking of perhaps a scenario with an Eclipse project for each core, customer, and group and then use Project Sets to select those we need for a given scenario. The problem I cannot get my head about, is how we would create robust code in the CORE projects which will work regardless of which group and customer is selected. A Factory class which knows which sub class of a passed Class object to invoke instead of each and every new?

Others must have had similar code base management problems. Anybody with experiences to share?

EDIT: The conclusion to this problem above has been that CVS needs to be replaced with a source code management system better suited for dealing with many branches concurrently and the migration of source from one component to the other while keeping history. Inspired by the recent migration by slf4j and logback we are currently looking at git as it handles branches very well. We've considered subversion and mercurial too but git appears to be better for single location, multibranched projects. I've asked about Perforce in another question, but my personal inclination is towards open source solutions for something as crucial as this.

EDIT: After some more pondering, we've found that our actual pain point is that we use branches in CVS, and that branches in CVS are the easiest to work with if you branch ALL files! The revised conclusion is that we can do this with CVS alone, by switching to a forest of java projects, each corresponding to one of the levels above, and use the Eclipse build paths to tie them together so each CUSTOMER version pulls in the appropriate GROUP and CORE project. We still want to switch to a better versioning system but this is so important a decision so we want to delay it as much as possible.

EDIT: I now have a proof-of-concept implementation of the CORE-GROUP-CUSTOMER concept using Google Guice 2.0 - the @ImplementedBy tag is just what we need. I wonder what everybody else does? Using if's all over the place?

EDIT: Now I also need this functionality for web applications. Guice was until the JSR-330 is in place. Anybody with versioning experience?

EDIT: JSR-330/299 is now in place with the JEE6 reference implementation Weld based on JBoss Seam and I have reimplemented the proof-of-concept with Weld and can see that if we use @Alternative along with ... in beans.xml we can get the behaviour we desire. I.e. provide a new implementation for a given functionality in CORE without changing a bit in the CORE jars. Initial reading up on the Servlet 3.0 specification indicates that it may support the same functionality for web application resources (not code). We will now do initial testing on the real application.