Consider setting ipvs sysctls

[SEJeff]

Per this redhat article, ipvs doesn't failover properly with udp services until the timeout is hit. This results in the data being blackholed for the duration of the timeout (defaults to 5 minutes). Setting this sysctl fixes this:

net.ipv4.vs.expire_nodest_conn: 1

Also from the article, this sysctl allows the backend weight to be set to 0 to quiesce the service (existing connections are not dropped but don't accept new connections):

net.ipv4.vs.expire_quiescent_template: 1

[SEJeff]

This article also has some extra background on the problem.

[roffe]

Running this in my staging cluster at the moment, image is avail at roffe/kube-router-git:issue-413

[SEJeff]

Awesome stuff thanks so much @roffe !