Please use proper version ranges when generating poms to be deployed to maven central #119
Comments
|
This will be challenging. It's already new to the most recent release that package requirements are mapped to maven dependencies in the first place, by the last rule here:
To implement that, we must resolve the package import to a bundle to get the coordinates of the dependency. But there is no information specified everywhere else about what might be an appropriate version range on that bundle. In in the case you're mentioning, you appear to want to exclude the next minor version, but it's far more common to only exclude the next major version expecting APIs to remain compatible. We'll need to add additional rules to covers special cases, but @mickaelistria has been trying to eliminate rules... |
|
I am working currently on https://issues.apache.org/jira/browse/MENFORCER-427 which allows to easily ban version which require resolving (version ranges and placeholders |
|
Maybe https://www.mojohaus.org/versions-maven-plugin/resolve-ranges-mojo.html can be added to the Maven build to resolve the versions only once during build. |
|
FWIW, we don't need extra plugins to achieve that; we do already know the compatible versions of dependencies when generating the pom files. The pom generator should just be changed to use those versions directly instead of generating a range from the requirement. |
|
@merks I just wanted to let you know that Tycho now includes a new extension component |
|
I have a feeling that there is something that needs to be clearly stated and properly understood here. For this specific example this IU has the following package requirements:
In general, these might or might not include a version range; in this case they have a one with only a lower bound. In any case, we must resolve each requirement to the corresponding capability/capabilities and then we decide that the IU(s) that provides those capabilities are to be treated as dependencies; this is new and was not supported before the most recent release. Keep in mind too that in general there is no relation between the IU's version and the capability's (package's) version so even if the package requirement has a range, that range cannot necessarily be mapped to a range of IUs. Even in this specific example, OSGi suggests that any package version higher than 1.69.0 is fine, but you're saying that's not correct either... In any case, yes, we could create a range for such a dependency based on the resolved IU available at the time the mapping is built; we already do that work just to get the IU needed for the dependency. But now the question is, is the very narrow range you are suggesting the only "proper" thing to produce? I think different users of the aggregator will have different ideas of what's proper and desired. It is after all extremely common to create ranges that exclude the next major version and we also see that it's problematic to consume 3rd party dependencies that have narrow dependency version ranges because it makes it hard to use other newer things. So I'm more than a little doubtful that "proper" is well defined and that everyone will generally agree it should be done exactly a certain way, e.g, with a very narrow range. |
You map a package import (that might resolve to any version >= 1.69) to an artifact (that is the actual bundle) and this bundles (maven) version should then be used, not a range, because that is the item the aggregator was built with. Given a single version in maven means: Please use this version, but if (because of conflicts with other dependencies) someone else suggest a higher version this is fine. and that's the closest match we can get between maven<->osgi here, that open version ranges on packages/bundles are also bad is just another story... |
|
@laeubi Thanks for the clarification! So the idea is to use the version of the resolved bundle as the dependency's version which is strong hint to use that version but is also effectively like a lower bound... |
|
The Maven dependencies are just used for building and should use the lower bound of things which are deployed at run time in the OSGi container. They should indicate the minimum version of a dependency they require from an API perspective. |
|
@merks basically yes, but keep in mind that OSGi Version is not always maven version of a given artifact it is just often the case. |
In particular, support the ability to specify the version range of the dependency, including expressions to extract the version components from the bundle that provides the capability that satisfies the requirement, e.g., [major.minor.micro,major+1.0.0) #11 eclipse-platform/eclipse.platform.releng#119
Use 4.26-I-builds in the SDK4Mvn.aggr. Eliminate the mapping for bouncycastle. eclipse-platform#119
|
I modified the SDK4Mvn.aggr to use the new support described here: eclipse-cbi/p2repo-aggregator#11 (comment) Unfortunately these builds never seem to be fully successful: |
|
@merks I'm not sure if I understand correctly
So was the code changed but this leads to failing builds, or are the failing builds unrelated but currently prevent one from testing it? |
|
This part, which uses the aggregator, works: https://ci.eclipse.org/releng/view/Publish to Maven/job/CBIaggregator/lastBuild/console These downstream triggered jobs fail quite a lot but sometimes succeed: https://ci.eclipse.org/releng/view/Publish%20to%20Maven/job/PublishPDEToMaven/ This one hasn't worked for several weeks: https://ci.eclipse.org/releng/job/PublishPlatformToMaven/ There are clues about the failures in files like this: Nothing changed in the aggregator since it was successfully used to publish the 4.25 release, until today. That why I asked on this issue which build is involved in producing the "missing" results: If it relies on the following build succeeding then one might have to wait forever: |
|
Actually what I said above I think is not entirely true. I think the aggregator did change after the release. I need to retest how it is handling snapshot versions because that's poorly tested I think... |
Otherwise building the -SNAPSHOT version gets messed up. #11 eclipse-platform/eclipse.platform.releng#119
|
In here: I see an error like this: But building the aggregation locally I see the jar exists:
|
|
Topic continues and evolves as #127 |
As already noticed in the recent IP issue, Platform currently produces (from maven point of view) very bad meta-data by using way to open version ranges. This can easily break consumers of such artifacts and threaten the stability of builds!
One example is:
https://mvnrepository.com/artifact/org.eclipse.platform/org.eclipse.equinox.p2.repository/2.6.200 which even use completely open ranges. That means maven will always resolve to the latest version on central, what obviously will only be a matter of time to break consumers.
This already broke Tycho here:
and even on a frozen branch here:
because the mentioned artifact now pulls in https://repo1.maven.org/maven2/org/bouncycastle/bcpg-jdk18on/1.72/ that was just release a few hours ago, while it was as far as I know compiled against https://mvnrepository.com/artifact/org.bouncycastle/bcpg-jdk18on/1.71
So if it is not know that a certain dependency can work (and compile!) with a range of version, the version used to produce that artifact should be used instead, what still will allow maven to use never version but prefer the given one if no one requests to do so see:
https://cwiki.apache.org/confluence/display/MAVENOLD/Dependency+Mediation+and+Conflict+Resolution#DependencyMediationandConflictResolution-DependencyVersionRanges
The text was updated successfully, but these errors were encountered: