Kubernetes Networking Issues After Upgrading to 1068.9.0

[Jeremy Derr]

I've allowed one of my test clusters to upgrade to 1068.9.0 from 1010.6.0. After doing this, docker networking does not mesh up with the kubernetes networking layer. My pods get conflicting 172.17.0.0/24 IP addresses, preventing most of them from accessing any network resources (generally, either the first or last to start with any specific IP, depending on circumstances), instead of the customary Kubernetes 10.244.0.0/16 addresses.

Restarting Docker, the Kubelet, the Kube Proxy, and lastly deleting and rescheduling all pods will (usually) resolve the issue.

I've noticed the same issue with the 2 beta channel servers I mix-in to my production cluster, and have thus had to remove them from service.

Between these two releases, has something changed in Docker's start process in CoreOS? The Docker version has not been revved.

[Brandon Philips]

What are you using for the kubernetes networking?

--
You received this message because you are subscribed to the Google Groups "CoreOS User" group.
To unsubscribe from this group and stop receiving emails from it, send an email to coreos-user...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Jeremy Derr]

Flannel. I am using a heavily updated version of the old CloudFormation templates that originally came in the Kubernetes codebase, complete with the early-docker systemd unit.

Restarting docker sometimes (but not always) seems to resolve the issue. Currently, I just have flannel set to start with no additional configuration. Is this a situation where I need to make sure it starts after early-docker.service but before docker.service?

--

Jeremy Derr

jer...@derr.me

When in trouble or in doubt, run in circles, scream and shout.