Hello all
Note: this email mixes two issues: Java 9 support and support of
languages other than Java. I discuss those two issues together because
the proposed experiment tries to address them together. This email is
also reworded as a wiki page at [1].
Last April we had an email thread about "Folder structure for Java 9
project". After 4 months of thinking and doing some experiments, I came
to question whether the Maven standard directory layout is still suited
to the current development landscape. My issues are:
* Maven standard directory layout is incompatible with the
--module-path options of JDK 9 tools (javac, javadoc, jlink,
/etc./). It has been argued that those options are only useful for
building more than one module at once. But I want this capability
because:
o I'm not aware of other way to generate satisfying aggregated
javadoc (MJAVADOC-449).
o It concerns not only aggregated javadoc, but most JDK tools
(e.g. annotation processors).
o It simplifies the handling of some circular dependencies.
* When using more than one language (in my case Java, Python and
C/C++) as components of the same project, I find the
"src/<main_or_test>/<language>" directory order tedious to use because:
o The tests are not close to the classes they are testing (less
than if the order was "src/<language>/<main_or_test>"). They are
interleaved with directories of other languages, thus
distracting navigation with less relevant content (from the
perspective of target language).
o Maven directory layout assumes that all languages split their
files in the same set of directories ("main" and "test"). But
Python and other languages have their own tools and practices.
o If a language needs another directory (e.g. "doc"), creating
"src/doc/<language>" imposes to all languages in that module a
directory relevant to only one language.
o "src/main/resources" is Java resources in practice. It would
have been more obvious if the path was "src/java/resources".
o Is the "src" level really necessary? After 15 years of using
Maven in various projects, it still appears like noise to me.
* The /"one pom.xml = one module"/ paradigm is not always convenient
because:
o If transposed to Python where a module is basically a single
file, it creates at least 4 files or directories boilerplate for
each Python file.
o It does not suit well to the build of many modules at once with
JDK 9 --module-path option; still workable, but not natural.
I realize that the /"one pom.xml = one module"/ paradigm is deeply
anchored in Maven architecture, but I came to a point where alternatives
seem yet more complicated (at least as a user). A directory layout
solving the above issues could be:
MyProject
├─ pom.xml
├─ MavenModule1
│ ├─ pom.xml
│ ├─ java
│ │ ├─ org.foo.bar1
│ │ │ ├─ main
│ │ │ │ └─ org/foo/bar1/…
│ │ │ └─ test
│ │ │ └─ org/foo/bar1/…
│ │ ├─ org.foo.bar2
│ │ │ ├─ main
│ │ │ │ └─ org/foo/bar2/…
│ │ │ └─ test
│ │ │ └─ org/foo/bar2/…
│ │ └─ org.foo.bar3
│ └─ python
│ ├─ foo
│ │ ├─ bar1.py
│ │ └─ bar2.py
│ └─ test_bar1.py
└─ MavenModule2
├─ pom.xml
...
A controversial point would be that a single Maven module, say
MavenModule1, could contain an arbitrary amount of Java 9 or Python
modules. The pom.xml for each Java 9 modules would either be inferred
automatically by computing the "intersection" of MavenModule1/pom.xml
with dependencies declared in the module-info of each Java module, or
specified explicitly in another way. GroupID and artifactID can be
inferred automatically too. After install or deploy phase, the JAR +
pom.xml files would be handled as today's Maven artifact.
This would require modifications of at least the Maven compiler,
javadoc, jar and deploy plugins (I realize that it would impact all
plugins, but other plugins could throw some "unsupported directory
layout" exception for now). I'm willing to experiment such layout and
Maven plugins modifications in the following months, because after 4
months of thinking I found no other satisfying solution except switching
to another build system. But comments are appreciated, and I would be
glad to propose the results back to Maven if the community consider them
acceptable.
Any though?
Martin
[1] https://github.com/Geomatys/maven-plugins/wiki
