Looking at https://github.com/apache/karaf/blob/e1dde27b88303c78e1d94b3dddaca0ea177e59e5/itests/test/pom.xml#L297-L313 it is clear this is not something solved in upstream and we will need to brew our own solution.

In order to control when a Karaf instance is launched and when not, we will need to attach to Maven reactor, hence this will invariantly include a bespoke maven plugin.

For maintaining reactor-global state we need to do something along the lines of https://github.com/fabric8io/docker-maven-plugin/issues/504 which ends up needing to talk to https://maven.apache.org/ref/3.8.2/maven-core/apidocs/org/apache/maven/execution/MavenSession.html .

That in turn gives us complete visibility into what the entirety of build looks like, with things like getCurrentProject() (hence no need for depends-maven-plugin), getProjects() and getPluginContext() – which should give us enough rope to implement what amounts to a Reactor-global lock we need to ensure there is no reactor-level concurrency around launching Karaf.

For launching Karaf, that does not look like an involved operation: https://github.com/ops4j/org.ops4j.pax.exam2/blob/master/containers/pax-exam-container-karaf/src/main/java/org/ops4j/pax/exam/karaf/container/internal/KarafTestContainer.java#L129 , hence we should be able to brew our own.

All we really need for SFT is to install the needed features and then monitor Karaf and wait for it to settle. I thing all the things we collect in BundleDiagInfosImpl should also be available through Karaf's serviceability features. Certainly the JMX information provided in osgi.core, org.apache.aries.blueprint and org.apache.karaf looks very promising.

As a side note: policy as to whether karaf can be launched concurrently or not depends on networking ports being used.
The first step there would be to reign in opendaylight-karaf-empty's, so that most of odlparent-packaged features can be tested concurrently – which would be an opt-in configured by the SFT-using artifact (based on knowledge of what is being tested). Based on that experience we may provide enough tools for at least some downstreams to the same.

That work is out of scope for this issue, though: for now we will serialize all Karaf tests.