How do I set up linkage between Docker containers so that restarting won't break it?

The effect of --link is static, so it will not work for your scenario (there is currently no re-linking, although you can remove links).

We have been using two different approaches at dockerize.it to solve this, without links or ambassadors (although you could add ambassadors too).

1) Use dynamic DNS

The general idea is that you specify a single name for your database (or any other service) and update a short-lived DNS server with the actual IP as you start and stop containers.

We started with SkyDock. It works with two docker containers, the DNS server and a monitor that keeps it updated automatically. Later we moved to something more custom using Consul (also using a dockerized version: docker-consul).

An evolution of this (which we haven't tried) would be to setup etcd or similar and use its custom API to learn the IPs and ports. The software should support dynamic reconfiguration too.

2) Use the docker bridge ip

When exposing the container ports you can just bind them to the docker0 bridge, which has (or can have) a well known address.

When replacing a container with a new version, just make the new container publish the same port on the same IP.

This is simpler but also more limited. You might have port conflicts if you run similar software (for instance, two containers can not listen on the 3306 port on the docker0 bridge), etcétera… so our current favorite is option 1.

Links are for a specific container, not based on the name of a container. So the moment you remove a container, the link is disconnected and the new container (even with the same name) will not automatically take its place.

The new networking feature allows you to connect to containers by their name, so if you create a new network, any container connected to that network can reach other containers by their name. Example:

1) Create new network

$ docker network create <network-name>       

2) Connect containers to network

$ docker run --net=<network-name> ...

or

$ docker network connect <network-name> <container-name>

3) Ping container by name

docker exec -ti <container-name-A> ping <container-name-B> 

64 bytes from c1 (172.18.0.4): icmp_seq=1 ttl=64 time=0.137 ms
64 bytes from c1 (172.18.0.4): icmp_seq=2 ttl=64 time=0.073 ms
64 bytes from c1 (172.18.0.4): icmp_seq=3 ttl=64 time=0.074 ms
64 bytes from c1 (172.18.0.4): icmp_seq=4 ttl=64 time=0.074 ms

See this section of the documentation;

Note: Unlike legacy links the new networking will not create environment variables, nor share environment variables with other containers.

This feature currently doesn't support aliases

You can use an ambassador container. But do not link the ambassador container to your client, since this creates the same problem as above. Instead, use the exposed port of the ambassador container on the docker host (typically 172.17.42.1). Example:

postgres volume:

$ docker run --name PGDATA -v /data/pgdata/data:/data -v /data/pgdata/log:/var/log/postgresql phusion/baseimage:0.9.10 true

postgres-container:

$ docker run -d --name postgres --volumes-from PGDATA -e USER=postgres -e PASS='postgres' paintedfox/postgresql

ambassador-container for postgres:

$ docker run -d --name pg_ambassador --link postgres:postgres -p 5432:5432 ctlc/ambassador

Now you can start a postgresql client container without linking the ambassador container and access postgresql on the gateway host (typically 172.17.42.1):

$ docker run --rm -t -i paintedfox/postgresql /bin/bash
root@b94251eac8be:/# PGHOST=$(netstat -nr | grep '^0\.0\.0\.0 ' | awk '{print $2}')
root@b94251eac8be:/# echo $PGHOST
172.17.42.1
root@b94251eac8be:/#
root@b94251eac8be:/# psql -h $PGHOST --user postgres
Password for user postgres: 
psql (9.3.4)
SSL connection (cipher: DHE-RSA-AES256-SHA, bits: 256)
Type "help" for help.

postgres=#
postgres=# select 6*7 as answer;
 answer 
--------
     42
(1 row)

bpostgres=# 

Now you can restart the ambassador container whithout having to restart the client.