After building a Kubernetes cluster, I tried to migrate my ZNC installation from a stand-alone VPS to Kubernetes.
I created a
Service, and I now had a ZNC instance running with persistent storage (for configuration, logs and so forth).
The only problem was that my ZNC instance couldn't connect to the internet. That's kind of important for ZNC, which is an IRC bouncer, and literally has only two jobs:
- Accept connections from IRC clients.
- Open connections to IRC servers.
After some quick debugging with
kubectl exec -it, I discovered that the Pod/container did actually have access to the internet, because I could ping public IP addresses. (Thank you
The problem was in DNS.
(Original source unknown, sourced from /r/sysadmin.)
After a couple hours of further debugging, I discovered the following:
All of my Kubernetes servers/VMs are running Ubuntu Server 17.10. Ubuntu uses systemd, and whilst I don't know much about systemd, I do know that it is heavily derided for being the exact opposite of the do-one-thing-and-do-it-well UNIX tooling philosphy.
In this case, systemd has it's own DNS resolver called
systemd-resolved. This replaces the traditional resolver, and
/etc/resolv.confredirects applications to use this instead.
resolv.conflooks like this (below).
If you do run
systemd-resolve --status, it lists a whole bunch of stuff you probably don't care about, including what your real upstream DNS servers are.
# This file is managed by man:systemd-resolved(8). Do not edit. # # 127.0.0.53 is the systemd-resolved stub resolver. # run "systemd-resolve --status" to see details about the actual nameservers. nameserver 127.0.0.53
- Kubernetes has it's own DNS resolver too, running inside a pod named
kube-dns. Inside a pod,
resolv.conflooks like this:
nameserver 10.96.0.10 search default.svc.cluster.local svc.cluster.local cluster.local options ndots:5
Kubernetes configures it's own DNS server to use the host's
/etc/resolv.confas an upstream for anything that it cannot resolve.
127.0.0.53is part of the
127.0.0.0/8range, it is the same as
Thanks to the default kernel namespaces provided by Docker/Kubernetes,
localhostinside of a Pod/container is not the same as
localhoston the host machine.
Of all of this, (5) is probably the most critical. If we zoom out to have a look at what happens when we make a DNS request:
- The pod looks at
- The pod makes a DNS request to
10.96.0.10, which is the
kube-dnscan't handle the request, so it sends it to an upstream DNS server,
kube-dnsis running inside a kernel networking namespace, it doesn't have access to the real
127.0.0.53:53address on the host. It's DNS requests stay within the pod, and are never answered.
The solution to this is to bypass
systemd-resolverd and get
kube-dns to talk to a DNS server that it can actually reach.
By default, Kubernetes supplies
kube-dns with upstream DNS server addresses from the host's
/etc/resolv.conf, however this is configurable. We can change these addresses by providing a
ConfigMap for the
Here's a simple one that just uses the primary Google DNS resolver:
apiVersion: v1 kind: ConfigMap metadata: name: kube-dns namespace: kube-system data: upstreamNameservers: | ["126.96.36.199"]
If you install this with
kubectl apply -f myconfig.yaml, then
kube-dns will be able to actually forward recursive DNS requests upstream. Your pods should now be able to happily resolve names to make connections to the Internet.
I actually suspect that it does an infinitely recursive lookup where it attempts to answer it's own upstream requests, because the
kube-dnspod was regularly crashing when I attempted to perform DNS resolution. ↩︎
You may also have to restart the