Vivid init system bug status

[Martin Packman]

Update on where we're at with the local provider init system breakage on vivid.

Oleg reported the following bug againsts 1.23:

<https://bugs.launchpad.net/juju-core/+bug/1443440>

Short version, juju sometimes fails when trying to operate on the
mongo service, during either bootstrap or destroy-environment. The
error is trying to run an upstart file, when the vivid system doesn't
have it because instead it's using systemd:

   fork/exec /sbin/initctl: no such file or directory

Theory of the bug, the code is looking for an error when running the
script, but gets a different error than expected when the script does
not exist. The randomness isu due to a loop over a map of init
systems, if we happen to try systemd before upstart all is good,
otherwise it blows up.

Based on this, Menno landed the following fix before 1.23 to fix the issue:

<https://github.com/juju/juju/pull/2083>

The problem is... somehow we can still hit the issue.

Oleg has done this two ways:

* Downloading the release tarball from launchpad, then building and
packaging locally and testing
* Installing the ppa:juju/proposed packages and testing

In addition, *one* of the CI test runs on the rev that included this
fix hit the problem:

<http://reports.vapour.ws/releases/2543/job/local-deploy-vivid-amd64/attempt/426>

Unfortunately we can't reproduce the failure otherwise on our vivid slaves.

Further, on a vivid machine in canonistack Oleg created, with minimal
setup, we could hit the issue often with our packaged juju versions,
but not go binaries built by Menno or I with added instrumentation.

So, as far as we can tell the bug should be fixed, somehow is not, but
we really do seem to be using the correct version of the source which
includes that code change.

Martin