For every module in a maven project generates a version which is a hash code of the module sources and its dependency tree (in the spirit of the Nix package manager). A build extension can skip modules if artifact for the same version already exists.

The goal is to avoid rebuild of modules which are not changed. Without splitting into multiple repositories, manual version management and the instability of the SNAPSHOT versions.

Mostly oriented to speedup CI server builds of multi-module projects.

Usage

Pure Informational

Run this in the root directory of a multi-module maven project:

    mvn pro.avodonosov:hashver-maven-plugin:1.6:hashver

This produces file target/hashversions.properties with hashversions for all modules in the project. The hashversions are in the form {module own hash}.{dependency tree hash}. If you change something in the code you can observe how hashversions are changed accordingly.

Avoid Unnecessary Rebuilds

The functionality described above is already sufficient in some sense: with every build generate a hashver file, and when publishing the artifacts record in some storage their hashversions. On the next build generate hashvers again, consult the storage and list the modules whose new hashversions are absent in the storage.

Then build only those modules, using mvn -pl <the modules> -am.

This approach does not even require touching the project's pom.xml files. See example in the "projects-to-build" mojo description below.

However, for this you will need to bother about the storage for hashversions, create supporting scripts. Also, such a build will rebuild unchanged dependencies of changed modules. For example, if subModuleX is changed in a project with the following structure:

- moduleA
  - subModuleX
  - subModuleY
- moduleB
- moduleC

the mvn -pl :moduleA,:subModuleX -am will also rebuild unchnaged subModuleY because it is required for moduleA build.

Therefore, it is more optimal to actually use hashversions as artifact versions. That way no separate storage is needed, all information is stored in the artifact repository. And unchanged dependencies of changed modules are not rebuilt but simply fetched from the repository.

Preparation steps.

An example of the below steps done for the maven-wagon project: https://github.com/avodonosov/maven-wagon/commit/fec377e145cd39187f392905156e6840c7629b75

Give every module an explicit version using property [artifactId].version, like <version>${mymodule.version}</version>. Use this property expression everywhere you refer the module version (as a dependency of other modules, etc).
Create file versions.properties in the root directory of the project, which specifies normal (non hash) values for all the module version properties. (You can copy the target/hashversions.properties we generated earlier and change all hash versions to normal versions)
Add the extension provided by the hashver-maven-plugin to the .mvn/extensions.xml. In usual builds this extension reads the version.properties and defines all those properties as system properties. In "hashversions mode" it will read the target/hashversions.properties instead, and will skip build of modules whose artifacts exist already.
Include flatten-maven-plugin so that the final pom's published with your artifacts have the version property expressions resolved.

Now the build will work as before, using the "normal" versions we specified in the versions.properties.

Note, maven prints warnings like

    [WARNING] 'version' contains an expression but should be a constant.  ...

they can be ignored - the flatten-maven-plugin takes care of the problem motivating this warning.

Skip Unaffected Modules

When we want to utilize the hashversions functionality, first run the mojo to produce the target/hashversions.properties, and then use -DhashverMode in your next maven command.

    mvn pro.avodonosov:hashver-maven-plugin:1.6:hashver
    mvn package -DhashverMode

The mojo cannot be run in the same maven invocation as other goals, it has to be run separately as shown above, due to a kind of "chicken and egg" problem. The system properties for version expressions, to be interpolated into the project tree and be in effect for the other goals, have to be defined very early, right after maven session starts, and at this early point maven is unable to produce dependency graph we need to compute hashversions. If we compute hashversions after the project graph is built, it's too late to apply them.

In the hashver build mode we can consider the target/hashversions.properties the main build result, because only part of the project artifacts are produced on the build environment (the affected ones) and the rest are only referred in the hashversions.properties. Your build publishing and running scripts should take this into account, for example publish the produced artifacts and the hashversions.properties, and when rolling out this build to a server take the hashversions.properteis and fetch all the artifacts according to it.

Assumptions

We assume all the module sources are located in the src/ directory. For every module we only hash the pom.xml, the src/ directory and the optional extraHashData parameter.

Reference

The "hashver" mojo

    mvn pro.avodonosov:hashver-maven-plugin:1.6:hashver \
          [-DextraHashData=someBuildProperty] \
          [-DincludeGroupId]

extraHashData - Any value you want to include into the hash calculation. It may reflect some build properties other than sources.
includeGroupId - The property names in the target/hashversions.properties will include group ID. For example, org.apache.maven.wagon.wagon-http.version instead of simply wagon-http.version.

The build extension

The following properties are supported. Some of them can be passed through either system properties (sys) or project properties in pom.xml (prj), others only through system properties, because they are checked before the pom.xml is read.

hashverMode (sys) - If set, the extension works as if -DskipExistingArtifacts -DsysPropFiles=target/hashversions.properties were specified.
sysPropFiles (sys) - A comma separated list of property files to read into system properties. If file name is prefixed with opt: the file is optional, otherwise it's required. Relative file names are treated realtive to the project root directory. Example:
```
-DsysPropFiles=/a.properties,opt:b.properties,src/c.properties
```
The files are loaded in the order they are specified, so if the same property is specified in several files, the last file wins.

Default value: versions.properties
skipExistingArtifacts (sys, prj) - When specified tries to find an artifact for every project module, and if the artifact exists - removes the module from maven session, thus skipping its build.
existenceCheckMethods (sys, prj) - How the artifact existence check is performed. A comma separated list, with the following values supported:
- resolve - The default. Invokes standard maven artifact resolution process. The downside of this method is that it performs artifact download, even when not needed for the build (not a dependency of any changed module). The advantage is the most native integration with maven - it supports all protocols, all repository properties specified in all levels (proxies, passwords, etc).
- local - Only check the local repository.
- httpHead - Try performing HTTP HEAD for the artifact URL built against base URL of every enabled repo. Does not support proxies and authentication.
Example
```
    -DexistenceCheckMethods=local,httpHead
```

The "projects-to-build" mojo

Extends the "hashver" mojo behavior with comparing the computed hashversions against a database of previously successfully built module hashversions, and producing a list of modules absent in the database (modules affected by changes since the successful builds).

The produced module list is saved to target/hashver-projects-to-build in the format suitable for the -pl (--projects) maven option, with intention to build only them:

    mvn install -pl "$(cat target/hashver-projects-to-build)" -am

The database of previously successfully build modules is a directory specified by property dbDir.

The mojo also produces content to be copied into the db directory if the build of those modules succeeds:

   cp -r targer/hashver-db-additions/*  the-db-directory/

Besides the dbDir property the mojo supports all the properties of the "hashver" mojo.

Example how this mojo is used in GitHub Actions workflow of the redisson project: https://github.com/avodonosov/redisson/commit/1d68c64d7dd370652dc303a76735af9998b01353

Design considerations

When only dependencies have changed, but the module own sources are not changed, strictly speaking, the module only needs to be re-tested, compilation could be skipped (unless a dependency instruments code or affects compilation otherwise). But we don't want to risk correctness by hunting this minor speedup, especially that Java compiler is very fast and most of the build time is usually spent on tests. So we just fully rebuild the module.

Discussion

Email thread with title "versioning by hashes to speedup multi-module build (a'la nix package manager)" on the Maven User List. https://lists.apache.org/thread.html/r0a4d687d9a6c315d0f90db59d7e6e7da5d71c18e94d6439c7a548dc2%40%3Cusers.maven.apache.org%3E

Maven older than 3.3.1

The .mvn/extensions.xml is only supported since maven 3.3.1 (released in 2015).

For older maven place the hashver maven plugin jar file to ${maven.home}/lib/ext.

Unfortunately, declaring the extension in the pom.xml doesn't work in our case, because our extension reads the property files to system properties in AbstractMavenLifecycleParticipant.afterSessionStart, which is not called when the extension is declared in the pom.xml.

https://maven.apache.org/examples/maven-3-lifecycle-extensions.html#use-your-extension-in-your-build-s

Issues

Report security issues at: https://hackerone.com/central-security-project/reports/new Other issues on github.

License

GNU AGPL v3.

avodonosov / hashver-maven-plugin