PR builds are inconsistent

[keithc-ca]

Both the build and test jobs in a PR test pipeline use the pull/nnn/merge reference: the problem is the meaning of that reference can change between when the build is launched and when the test portion begins. See #6822 where the build was based on 98a888ab596c4cec379318bd9ae192ae633f3a9b, but the test was based on 15afcbb98d2f683190b07ca803e189749b6c64c5 because in the interim, #6938 was merged which changed the meaning of pull/6822/merge. The result in [1] is errors that refer to lines that don't exist in the commit used to build.

[1] https://ci.eclipse.org/openj9/job/Test_openjdk8_j9_sanity.functional_ppc64le_linux_Personal/260

[keithc-ca]

fyi @AdamBrousseau

[AdamBrousseau]

May be able to just remove this line.
https://github.com/eclipse/openj9/blob/master/buildenv/jenkins/common/pipeline-functions.groovy#L403-L405
I can't why we started this override in the first place.

[keithc-ca]

Google lead me to [1] which suggested that we could use the SHA directly:

git fetch upstream 98a888ab596c4cec379318bd9ae192ae633f3a9b:refs/remotes/upstream/merge-6822

but github doesn't want to cooperate:

error: Server does not allow request for unadvertised object 98a888ab596c4cec379318bd9ae192ae633f3a9b

This seems to be a relatively rare problem so we may just have to deal with it when it arises; and now that we understand why it might occur, it's easier to accept that me may need to relaunch PR builds on occasion.

[1] https://stackoverflow.com/questions/14872486/retrieve-specific-commit-from-a-remote-git-repository

[AdamBrousseau]

Also found this page
https://discourse.drone.io/t/github-claims-that-merge-refs-are-undocumented-feature/1100/2
From 2017 where Github recommends creating your own merge refs instead of using theirs.

[keithc-ca]

It's not clear there are any practical solutions to this. It happens infrequently enough that I think we can just live with the occasional such problem - closing.