You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@maven.apache.org by Max Bowsher <ma...@ukf.net> on 2007/05/18 19:11:13 UTC
Artifact resolution bugs and unfortunate jar-plugin/assembly-plugin
interactions.
This is an email about a class of bugs, discovered working with the
following general scenario:
I am building a standalone Java application.
maven-jar-plugin is configured with mainClass and addClasspath=true.
maven-assembly-plugin is configured with a <dependencySet> to gather up
all the dependencies into a single directory.
The application is invoked by a shell script which executes "java -jar
$jarfile".
This all works fine when using only release dependencies. When snapshot
dependencies are involved, the jar and assembly plugins sometimes
disagree about whether to use SNAPSHOT or YYYYMMDD.HHMMSS-N naming for
the dependent artifacts.
To begin with, I will discuss the situation when using
maven-assembly-plugin 2.1:
This is JIRAed as MNG-2456. The underlying issue here is a Maven core
bug - maybe something to fix in 2.1? Specifically, the
DefaultArtifactResolver explictly set the Artifact "file" field to a
path in the local repository using the SNAPSHOT rather than
YYYYMMDD.HHMMSS-N form. Perhaps this is correct behaviour - but if so,
it should be documented in the Artifact interface, and perhaps an
additional field made available to expose the versioned file, if one exists.
Now, switch to considering the situation with maven-assembly-plugin
2.2-beta-1:
This opens the doors on a larger class of bugs. Before, both jar and
assembly plugins were using the same set of resolved artifacts - the one
produced by Maven in response to the @requiresDependencyResolution mojo
javadoc annotation. assembly-plugin 2.2-beta-1, though, switches to
doing its own resolution. (I don't know why.) Even the "baseVersion"
fields of the returned artifacts are then set to YYYYMMDD.HHMMSS-N form
- clearly wrong.
This is a symptom, I think, of the weird way the DefaultArtifact class
is written, with accessors that cause state changes - bug prone. A key
example is DefaultArtifact.isSnapshot(). I would suggest that for Maven
2.1, the class be changed such that all state changes are done in setter
methods, rather than adapting the state of the artifact object on the
fly in accessor methods.
Whilst speaking of DefaultArtifact and versioning bugs, another aspect
to consider changing is the somewhat confusing interface for setting
this information: There are three private fields closely concerned with
this - the "baseVersion", "version", and "file". Each of these has a
direct public setter - though the setter for "version" is called
setResolvedVersion(), not setVersion(). Then, there are additional
setter methods setVersion(), selectVersion(), and updateVersion() which
update various combinations of these fields. I'd imagine some of these
should be deprecated, because the current situation makes it really hard
to determine what the true behaviour of the class is.
So, in conclusion, I think changes of some sort are needed in the maven
core, the jar plugin, and the assembly plugin, to fully sort out this
collection of bugs for both 2.0.x and 2.1 lines.
This is a really entangled issue with fearsome potential for
compatibility problems, but if people more expert than I can help with
defining the general direction to proceed, I'm more than willing to work
on producing patches.
Thanks,
Max.
Re: Artifact resolution bugs and unfortunate jar-plugin/assembly-plugin interactions.
Posted by Brett Porter <br...@apache.org>.
Seems like a good idea to capture this as a 2.1 feature request,
write it up in the wiki and link in any related JIRA issues from all
the plugins and MNG itself so that it has a focal point. As we work
through things like the artifact mechanism it can be taken into
consideration.
That's in the general case, but hopefully we can also apply
workarounds to the plugins to get these working properly in the mean
time with 2.0.x.
I'm also particularly interested in seeing the "wider class of bugs"
found by upgrading the assembly plugin be followed through while it's
still in beta.
- Brett
On 19/05/2007, at 3:11 AM, Max Bowsher wrote:
> This is an email about a class of bugs, discovered working with the
> following general scenario:
>
> I am building a standalone Java application.
> maven-jar-plugin is configured with mainClass and addClasspath=true.
> maven-assembly-plugin is configured with a <dependencySet> to
> gather up
> all the dependencies into a single directory.
> The application is invoked by a shell script which executes "java -jar
> $jarfile".
>
> This all works fine when using only release dependencies. When
> snapshot
> dependencies are involved, the jar and assembly plugins sometimes
> disagree about whether to use SNAPSHOT or YYYYMMDD.HHMMSS-N naming for
> the dependent artifacts.
>
> To begin with, I will discuss the situation when using
> maven-assembly-plugin 2.1:
>
> This is JIRAed as MNG-2456. The underlying issue here is a Maven core
> bug - maybe something to fix in 2.1? Specifically, the
> DefaultArtifactResolver explictly set the Artifact "file" field to a
> path in the local repository using the SNAPSHOT rather than
> YYYYMMDD.HHMMSS-N form. Perhaps this is correct behaviour - but if so,
> it should be documented in the Artifact interface, and perhaps an
> additional field made available to expose the versioned file, if
> one exists.
>
> Now, switch to considering the situation with maven-assembly-plugin
> 2.2-beta-1:
>
> This opens the doors on a larger class of bugs. Before, both jar and
> assembly plugins were using the same set of resolved artifacts -
> the one
> produced by Maven in response to the @requiresDependencyResolution
> mojo
> javadoc annotation. assembly-plugin 2.2-beta-1, though, switches to
> doing its own resolution. (I don't know why.) Even the "baseVersion"
> fields of the returned artifacts are then set to YYYYMMDD.HHMMSS-N
> form
> - clearly wrong.
>
> This is a symptom, I think, of the weird way the DefaultArtifact class
> is written, with accessors that cause state changes - bug prone. A key
> example is DefaultArtifact.isSnapshot(). I would suggest that for
> Maven
> 2.1, the class be changed such that all state changes are done in
> setter
> methods, rather than adapting the state of the artifact object on the
> fly in accessor methods.
>
> Whilst speaking of DefaultArtifact and versioning bugs, another aspect
> to consider changing is the somewhat confusing interface for setting
> this information: There are three private fields closely concerned
> with
> this - the "baseVersion", "version", and "file". Each of these has a
> direct public setter - though the setter for "version" is called
> setResolvedVersion(), not setVersion(). Then, there are additional
> setter methods setVersion(), selectVersion(), and updateVersion()
> which
> update various combinations of these fields. I'd imagine some of these
> should be deprecated, because the current situation makes it really
> hard
> to determine what the true behaviour of the class is.
>
>
> So, in conclusion, I think changes of some sort are needed in the
> maven
> core, the jar plugin, and the assembly plugin, to fully sort out this
> collection of bugs for both 2.0.x and 2.1 lines.
>
> This is a really entangled issue with fearsome potential for
> compatibility problems, but if people more expert than I can help with
> defining the general direction to proceed, I'm more than willing to
> work
> on producing patches.
>
>
> Thanks,
> Max.
>
>
>
>
---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@maven.apache.org
For additional commands, e-mail: dev-help@maven.apache.org