Summary of Solution

We need:

  1. building an attachment (distribution primarily) must build the original, and deploy both.
  2. type to packaging + classifier mappings (but no need for the reverse)
  3. binding for the package step (eg ejb-client, tld) - this will also be applicable to deployment, snapshot or not
  4. binding for the release step (eg sources, javadoc, distributions, etc)

Building an attachment must build both

As the assembly is blended into the lifecycle when a release is being performed, the packaging is already assured.

Type -> packaging + classifier mapping

Plugins can specify implementations of type handlers (and many will be in the core). These will indicate the packaging and classifier that they belong to. The default implementation will use packaging = type, no classifier.

Attachment at packaging step

Here, a simple call to project.attachArtifact( Artifact artifact ) is appropriate. In the specific use cases:

Attachment at deploy step

We want attachment of sources and javadoc to be the default, but we also want to configure in the attachment of various assemblies.

This will also depend on whether it is a snapshot build, or a release build as to what is actually attached. Profiles should help here.

We will introduce a a package-additional step after package, where this will occur (and they will use project.attachArtifact to trigger installation and deployment). Sources and javadoc will be enabled by default for jar:jar by the jar packaging lifecycle bindings. These will currently just check for the existence of SNAPSHOT as to whether they will do anything, but in future they should always do their task, and only be bound when in release mode (unless the user has specifically bound them somehow?)

Original Discussion - Differences Between Packaging and Dependency Type


I'd like to work through any problems that might arise from the
differences between <packaging> and <type>. Below is the background,
problem statement, and design decision to rectify them. Please add any
additional problems you see, or issues with the design.


In Maven 2.0, projects are keyed uniquely on only groupId and artifactId
for considering what is a project (version does come into it for
determining different versions of a project).
A packaging is specified on pom.xml to signify how to build that
project, and what type of artifact it produces.

For a given packaging, it may produce multiple artifacts that share the
same project information - the main example used is an ejb that also
outputs an ejb-client. There are also some goals that will produce
artifacts from a given project - for example, the assembly/distribution
goals, javadoc and sources.

To be able to depend on only some of those artifacts, <type> on a
dependency may not match packaging - if you declare <type>ejb</type> you
only get the ejb, and if you declare <type>ejb-client</type>, you only
get the ejb-client JAR. The POM can always be located by the
groupId/artifactId/version combination (type is not present in the m2
repo, and the POM is always under /poms/ if using the old layout).

This also means that while type is not considered in whether an project
is unique, it is for a dependency - so you can download tld files with
the same name as the JAR they belong to.

This is pretty sound from general use cases, though there have been a
couple of problems to arise:


I believe this is ok. The naming was different to ensure it is clear
that they can be different, and packaging IS A SUBSET of <type>,
guaranteeing they can be handled consistently. Any value of packaging
should have 1 type that matches exactly, and possibly more types that
are associated with it. Any type has 0..1 matching packaging.


If a JAR is deployed, then a distribution is deployed - either the POM is
republished (so the snapshot is bumped up and the JAR is orphaned), or
it is not and the published dist is orphaned, or reuses the timestamp
which is not technically correct (especially as it may overwrite a
previous distribution).


Firstly, we differentiate types in the following way:

Untied objects can be manually attached, and tied objects can be manually detached. From this, it seems sensible that attachment is only a matter of configuration, with sensible defaults, rather than something specific in the design. Note that there are different scenarios: attachment by configuration is appropriate for ejb client, but when set it should always happen. For distributions, sources, etc it is appropriate to only attach at release - they do not need to be deployed for snapshots.

This ensures that a POM is always deployed and updated for every
deployment of anything, and that the primary artifact and anything tied
to it is never orphaned.

It can mean that under the default lifecycle, there will be no snapshots
deployed for a distribution, for example. This simply means you cannot
depend on a distribution where you can a JAR, unless it is tied, which
is reasonable.

This does not affect released artifacts, as they can be released at
different times and still have the same version.

Declaring the type in a dependency

It is envisaged that the type could encapsulate the extra information about the classifier. ie, ejb-client is a type that maps to the packaging ejb with the classifier "client".

However, there are a few interesting points here:

I am in favour of using type mappings, at least to start with. The positives are: