merge-proposal-jobs OOPS-1628MPJ1 in cancel_deletion when merging into empty tree

[Steve McInerney]

 KeyError: None

    Traceback (most recent call last):
  Module lp.services.job.runner, line 189, in runJobHandleError
    self.runJob(job)
  Module lp.services.job.runner, line 160, in runJob
    job.run()
  Module lp.code.model.branchmergeproposaljob, line 337, in run
    self.branch_merge_proposal)
  Module lp.code.model.diff, line 338, in fromBranchMergeProposal
    prerequisite_branch)
  Module lp.code.model.diff, line 118, in mergePreviewFromBranches
    merge_target, cleanups)
  Module lp.code.model.diff, line 149, in _getMergedTree
    transform = merger.make_preview_transform()
  Module bzrlib.merge, line 788, in make_preview_transform
  Module bzrlib.cleanup, line 122, in run_simple
  Module bzrlib.cleanup, line 156, in _do_with_cleanups
  Module bzrlib.merge, line 793, in _make_preview_transform
  Module bzrlib.merge, line 821, in _compute_transform
  Module bzrlib.merge, line 1044, in fix_root
  Module bzrlib.transform, line 323, in cancel_deletion
KeyError: None
https://lp-oops.canonical.com/oops.py/?oopsid=1628MPJ1

or, https://pastebin.canonical.com/33566/

<poolie> i wonder if this is bad exception cleanup code
<poolie> yes it probably is
<poolie> this is a knock on error from some other error (possibly ioerorr, or interrupt)

Erroring creating a preview tree for merging
  lp:~ubuntu-drupal-devs/ubuntu-website/light-drupal-theme-v6.6.0
into
  lp:~ubuntu-website-community/ubuntu-website/light-drupal-theme

[Robert Collins (lifeless)]

This needs someone to examine the merge-proposal-jobs code and exclude it from consideration, for the bzr task to be marked confirmed.

[Tim Penhey (thumper)]

I'm wondering what we should be doing in the code side of things.

Perhaps we need to capture the failure and retry the job again. Jobs have this facility, but we aren't really using it yet... perhaps we should be.

If we are going to retry, we should integrate into the base job class a max_retries.

[Aaron Bentley (abentley)]

I think we should reserve retries for situations where we have some reason to expect that it would succeed if we ran it again, e.g. a connectivity problem. I don't see any reason to believe that here.

[Aaron Bentley (abentley)]

The most recent instance, OOPS-1979MPJ1 appears to be due to attempting to perform a merge into the empty tree. While bzr may not do this, the failure is clearly a bug in bzrlib caused by the merge code assuming that the tree has a root.

[Данило Шеган (danilo)]

Is there anything we've got to do for LP for this bug to be resolved (other than potentially upgrade to latest bzr if we haven't already)?

[Aaron Bentley (abentley)]

Upgrading bzr was all that was required.