weavedns

Weave DNS server

The Weave DNS server answers name queries in a Weave network. This provides a simple way for containers to find each other: just give them hostnames and tell other containers to connect to those names. Unlike Docker 'links', this requires no code changes and works across hosts.

Using weaveDNS

WeaveDNS is deployed as a set of containers that communicate with each other over the weave network. One such container needs to be started on every weave host, by invoking the weave script command launch-dns. Application containers are then instructed to use WeaveDNS as their nameserver by supplying the --with-dns option when starting them. Giving any container a hostname in the .weave.local domain registers it in weaveDNS. For example:

$ weave launch
$ weave launch-dns 10.1.254.1/24
$ weave run 10.1.1.25/24 -ti -h pingme.weave.local ubuntu
$ shell1=$(weave run --with-dns 10.1.1.26/24 -ti -h ubuntu.weave.local ubuntu)
$ docker attach $shell1

# ping pingme
...

Each weaveDNS container started with launch-dns needs to be given its own, unique, IP address, in a subnet that is a) common to all weaveDNS containers, b) disjoint from the application subnets, and c) not in use on any of the hosts. In our example the weaveDNS address is in subnet 10.1.254.0/24 and the application containers are in subnet 10.1.1.0/24.

WeaveDNS containers can be stopped with stop-dns.

How it works

The weaveDNS container running on every host acts as the nameserver for containers on that host. It is told about hostnames for local containers by the weave run command. If a hostname is in the .weave.local domain then weaveDNS records the association of that name with the container's weave IP address(es).

When weaveDNS is queried for a name in the .weave.local domain, it first checks its own records. If the name is not found there, it asks the weaveDNS servers on the other hosts in the weave network.

When weaveDNS is queried for a name in a domain other than .weave.local, it queries the host's configured nameserver, which is the standard behaviour for Docker containers.

Domain search paths

If you don't supply a domain search path (with --dns-search=), weave run ... tells a container to look for "bare" hostnames, like pingme, in its own domain. That's why you can just invoke ping pingme above -- since the hostname is ubuntu.weave.local, it will look for pingme.weave.local.

If you want to supply other entries for the domain search path, e.g. if you want containers in different sub-domains to resolve hostnames across all sub-domains plus some external domains, you need also to supply the weave.local domain to retain the above behaviour.

weave run --with-dns 10.1.1.4/24 -ti \
  --dns-search=zone1.weave.local --dns-search=zone2.weave.local \
  --dns-search=corp1.com --dns-search=corp2.com \
  --dns-search=weave.local ubuntu

Doing things more manually

If you use the --with-dns option, weave run automatically supplies the DNS server address to the new container. And both weave run and weave attach register the hostname of the given container against the given weave network IP address.

In some circumstances, you may not want to use the weave command. You can still take advantage of a running weaveDNS, with some extra manual steps.

Using a different docker bridge

So that containers can connect to a stable and always routable IP address, weaveDNS publishes its port 53 to the Docker bridge device, which is assumed to be docker0.

Some configurations may use a different Docker bridge device. To supply a different bridge device, use the environment variable DOCKER_BRIDGE, e.g.,

$ sudo DOCKER_BRIDGE=someother weave launch-dns 10.1.254.1/24

Supplying the DNS server

If you want to start containers with docker run rather than weave run, you can supply the docker bridge IP as the --dns option to make it use weaveDNS:

$ docker_ip=$(docker inspect --format='{{ .NetworkSettings.Gateway }}' weavedns)
$ shell2=$(docker run --dns=$docker_ip -ti ubuntu)
$ weave attach 10.1.1.27/24 $shell2

This isn't very useful unless the container is also attached to the weave network (as in the last line above).

Also note that this means of finding the Docker bridge's IP address requires a running container (any one would do); another way to find it is:

$ docker_ip=$(ip -4 addr show dev docker0 | grep -o 'inet [0-9.]*' | cut -d ' ' -f 2)

Supplying the domain search path

By default, Docker provides containers with a /etc/resolv.conf that matches that for the host. In some circumstances, this may include a DNS search path, which will break the nice "bare names resolve" property above.

Therefore, when starting containers with docker run instead of weave run, you will usually want to supply a domain search path so that you can use unqualified hostnames. Use --dns-search=. to make the resolver use the container's domain, or e.g., --dns-search=weave.local to make it look in weave.local.

Adding containers to DNS

If DNS is started after you've attached a container to the weave network, or you want to give the container a name in DNS other than its hostname, you can register it using the HTTP API:

$ docker start $shell2
$ curl -X PUT "http://$dns_ip:6785/name/$shell2/10.1.1.27" -d fqdn=shell2.weave.local

Registering multiple containers with the same name

This is supported; weaveDNS picks one address to return when you ask for the name. Since weaveDNS removes any container that dies, this is a simple way to implement redundancy. In the current implementation it does not attempt to do load-balancing.

Replacing one container with another at the same name

If you would like to deploy a new version of a service, keep the old one running because it has active connections but make all new requests go to the new version, then you can simply start the new server container and then unregister the old one from DNS. And finally, when all connections to the old server have terminated, stop the container as normal.

Not watching docker events

By default, weaveDNS watchs docker events and removes entries for any containers that die. You can tell it not to, by adding --watch=false to the container args:

$ weave launch-dns 10.1.254.1/24 --watch=false

Unregistering

You can manually delete entries for a host, by poking weaveDNS's HTTP API with e.g., curl:

$ docker stop $shell2
$ dns_ip=$(docker inspect --format='{{ .NetworkSettings.IPAddress }}' weavedns)
$ curl -X DELETE "http://$dns_ip:6785/name/$shell2/10.1.1.27"

Present limitations

• The server will not know about restarted containers, but if you re-attach a restarted container to the weave network, it will be re-registered with weaveDNS.
• The server may give unreachable IPs as answers, since it doesn't try to filter by reachability. If you use subnets, align your hostnames with the subnets.
• We use UDP multicast to find out about remote names (from weaveDNS servers on other hosts); this likely won't scale well beyond a certain point T.B.D., so we'll have to come up with another scheme.