We practice highly componentised development using Java, we have about 250 modules in trunk that have independent life cycles. Dependencies are managed through Maven (that’s a best practice right there), every iteration (bi-weekly) actively developed modules get tagged with a new version. 3 digit version numbers with strict semantics (major.minor.build – major changes means backwards incompatible, minor changes mean backwards compatible and build number changes mean backwards and forwards compatible). Our ultimate software product is an assembly that pulls in dozens of individual modules, again as Maven dependencies.
We branch modules/assemblies when we need to make a bug fix or enhancement for a released version and we can not deliver the HEAD version. Having tagged all versions makes this easy to do but branches still incur a significant administrative overhead (specifically keeping branches in sync with certain HEAD changesets) that are partly caused by our tools, Subversion is sub-optimal for managing branches.
We find that a fairly flat and above all predictable tree structure in the repository is crucial. It has allowed us to build release tools that take away a lot of the pain and danger from a manual release process (updated release notes, project compiles, unit tests run through, tag is made, no SNAPSHOT dependencies, etc). Avoid putting too much categorization or other logic in your tree structure.
We roughly do something like the following:
svnrepo/
trunk/
modules/
m1/ --> will result in jar file
m2/
...
assemblies/
a1/
...
tags/
modules/
m1/
1.0.0/
1.0.1/
1.1.0/
m2/
...
assemblies/
a1/
iteration-55/
...
branches/
m1/
1.0/
...
For external dependencies, I can not overemphasize something like Maven: manage your dependencies as references to versioned, uniquely identified binary artifacts in a repository.
For intenal module/project structure: stick to a standard. Uniformity is key. Again, Maven can help here since it dictates a structure. Many structures are fine, as long as you stick to them.