When I started the mavenization of my projects, in July, I really didn't figure out that it would have been such a painful and long process. It is literally consuming me - also because I'm longing to see the end of the conversion, so I can resume the development.

The first mavenized projects, BetterBeansBinding, jrawio and Mistral, were reasonably easy, also considering that I had to learn at the same time Maven under the surface. But they were plain JSE projects, with just a handful of modules. blueMarine has got hundreds of modules partitioned in 7 projects and even though I wrote some scripts and tools to automatically generate the pom files, they are needing tons of manual fixing.

So you can imagine my enthusiasm when, last night, the first subproject of blueMarine (Metadata) at last compiled correctly using only Maven, so I could drop all the Ant artifacts. Hopefully the second subproject should be converted in the weekend, so I'll be able to start blogging about this topic: mavenizing a NetBeans Platform project opens a number of issues, from the management of dependencies (e.g. transitive vs non transitive dependencies) to the lifecycle of versions.

As a first comparison, I've benchmarked how long it takes to build the project (in my optimized build environment):

Clean and build with existing build artifacts
ant clean nbms	26"
mvn clean install -DskipTests	1'35"
Build from a clean status (no platform, no build artifacts)
ant nbms	24"
mvn install -DskipTests	1'32"
Build from a clean status with existing platform (no build artifacts)
ant nbms	22"
mvn install -DskipTests	1'32"

You can get the source code with hg clone https://kenai.com/hg/bluemarine~metadata-src and use revision d52dc0fff7d3 for benchmarking Maven and 79f2f26dfbe7 for benchmarking Ant. Of course all the tests have been performed with all the needed files already on the disk, so there were no network downloads involved.

As expected, Maven is a clear loser here, with roughly a 3.5x factor in build times. Part of the problem is due to the unflexible nature of Maven: for instance, an install always rebuilds the platform from the nbm artifacts in the local repository, while my Ant scripts were smarter and performed a check to execute that operation only when needed. Another obvious point looking at the console is that, in spite of tests being disabled, Maven still runs Cobertura, which instruments the code before realizing that there are no tests to run. It's the usual Maven dumbness.

Hopefully, there is some room for improvements here, for instance by using profiles and disabling some plugins when they aren't needed. Matter for the next post.