aws: workaround for systemd #1312

[mumoshu]

On top of #682 (comment), replaces a dependency from kubelet to the oneshot service decrypt-tls-assets with an ExecStartPre in the kubelet service because systemd doesn't seem to restart failed decrypt-tls-assets services that way.

Also, I've changed where to add ExecStartPre, instead of [Unit] sections as done in https://gist.github.com/spacepluk/a14f10cfed3756c0f1f079e73cdc6c9a#gistcomment-1885799, to [Service] sections as done in this PR.

Mainly purposed to fix #675 which I've encountered earlier, but would fix #682 too.

Although my cluster brought up via kube-aws with this change is working without any problem,
I have not yet a chance to reproduce #682 to test this.
Is anyone aware of a handy way to reproduce it?

cc @cgag @spacepluk

[dghubble]

I think this takes the systemctl is-active some.service trick/hack too far. That only works if the service itself has restarts, otherwise some of these changes "hope" the dependency is satisfied and fail if not. I don't think its useful for the oneshot for example. Requires= or Wants= are still the right way to do this.

The is-active flanneld was done on clusters which required a remote etcd to come online (via polling) before flannel starts and thus kubelet can start if I recall correctly. Starting etcd -> flannel -> kubelet should work here since flannel just depends on the local system's etcd.

• restarting systemd service on dependency failure systemd/systemd#1312
• https://github.com/coreos/coreos-baremetal/issues/253

[mumoshu]

@dghubble Thanks again for your feedback!

I guess things became a bit clearer for me now.

To sum up,

• Modification to load-rkt-stage1.service seems meaningless because it is a oneshot service and won't try to restart when ExecStartPre failed, by its nature
• Can network-online.target fail transiently? Maybe no. Then, modifications to calico-node.service seems to be meaningless in practice.
• Modification to kubelet.service would be useful according to coreos-baremetal#253
• Modification to flanneld.service would be useful according to controller node fails to start periodically because flannel doesn't start. #710
• Modification to docker.service would be useful according to why is docker.service not automatically restarting #682

[aaronlevy]

I'm going to close this PR, as the kube-aws work has moved to its own top-level repository here: https://github.com/coreos/kube-aws . If this is still something you would like to be merged, please open a PR under the new repo.