We've had issues with KubeDNS, too. Lots of retries and timeouts on the client side, and lots of conntrack entries. Libc has pretty slow retries (5s, I think)

We've had issues with KubeDNS, too. Lots of retries and timeouts on the client side, and lots of conntrack entries. Libc has pretty slow retries (5s, I think) by default, and until 1.11 hits you can't easily set up resolver configs, though you can inject an envvar separately into each. And musl-based distros like Alpine don't even support some of libc's options, iirc. We ended up scaling up KubeDNS to 2 replicas and moving them to a dedicated nodepool just to make sure they weren't competing with other nodes. That fixed our issues for now.