Attempt to report agent failures as INFRA

[johnsca]

Recently, we've had a spate of failures on GCE. For some reason, they report as hook failures, even though it seems that what actually happened is that the agent crashed. For example: http://cwr.vapour.ws/bundle_hadoop_kafka/361b12d5cace479a8ed3a9965e4610c7/report.html

We should look into whether it's possible to treat this as INFRA. Specifically, if "machine" is anything other than "started" is probably an INFRA.

[kwmonroe]

Since the failures are so widespread and recent for gce, this may be related:

http://reports.vapour.ws/releases/issue/57c3c12c749a564be035e4ee

This is an issue where the controller can't contact agents, and there are numerous reports of this happening for gce over the last week.

[kwmonroe]

Is this issue still valid? I haven't seen any hook failures due to juju agent crashing recently. I have seen INFRA failures for things like "gce cannot upload scripts", but that's not the same.

Perhaps this fixed itself with the move from juju-deployer (v1) to juju deploy (v2).

[johnsca]

I think it would still be reasonable to detect the machine agent status and report that as INFRA even if workload-status is error.