DebugMetricsCommandSuite.TearDownTest fails due to "no reachable servers."

[Christopher Lee]

As seen at: http://reports.vapour.ws/releases/issue/57bd1b8d749a5638f818a04e

The unit test teardown (DebugMetricsCommandSuite.TearDownTest) fails with:

failed: no reachable servers")
... error stack:
 no reachable servers
 github.com/juju/testing/mgo.go:632: inst.Dial() failed

Which is potentially caused by a socket connection refusal:
[LOG] 0:03.650 DEBUG juju.mongo connection failed, will retry: dial tcp 127.0.0.1:43893: getsockopt: connection refused
setmeterstatus_test.go:105:
    c.Assert(err, jc.ErrorIsNil)
... value *errors.Err = &errors.Err{message:"cannot retrieve meter status for metrics manager", cause:(*errors.errorString)(0xc82000c110), previous:(*errors.Err)(0xc820492280), file:"github.com/juju/juju/state/meterstatus.go", line:150} ("cannot retrieve meter status for metrics manager: EOF")
... error stack:
 EOF
 github.com/juju/juju/state/errors.go:57:
 github.com/juju/juju/state/meterstatus.go:150: cannot retrieve meter status for metrics manager

[Casey Marshall (cmars)]

I don't think this is anything to do with the test itself. I think it's a general flakiness that I've observed in MgoTestSuite lately. I see lots of connection refused errors in CI test attempts. And, they seem to occur in different tests.

I wonder if it'd be helpful for testing/mgo to keep track of the mongod logs and dump them when dialing fails like this -- especially in setup/teardown, to help us understand & fix the problem.