Grouping applications
Juju deploys units of an application from a charm. The simplest way to do this is
juju deploy mysql
which will take the latest version of the MySQL charm straight from the store and create a single application unit.
By default, Juju creates a unit of an application named after the charm; mysql 
in the above example.
You can specify the name for the application when deploying:
juju deploy mysql wikidb
which will create a unit of the wikidb application. It creates the unit 
exactly as before, except instead of mysql, the application is called 
wikidb.
Why specify names for applications? The simplest reason is organizational. It lets you stay organized as your infrastructure gets more complex:
juju deploy mysql website-db juju deploy mysql app-master-db juju deploy mysql app-slave-db -n2
But there are other reasons to do this: application groups, which are nothing other than named applications. They are called out as a separate feature because they're quite useful in a few different ways.
Roles
Some applications acquire a role at runtime based on the relations that are joined.
A great example of this is the hadoop charm. It can instantiate applications
juju deploy hadoop namenode juju deploy hadoop datacluster -n40
which are identical at this point except for the application name that Juju's using for the various units.
These applications acquire roles at relation-time via
juju add-relation namenode:namenode datacluster:datanode juju add-relation namenode:jobtracker datacluster:tasktracker
The relations determine the application role.
Another example of this is mysql replication.
juju deploy mysql masterdb juju deploy mysql slavedb -n2
where the different applications are related to each other
juju add-relation masterdb:master slavedb:slave
and to other applications via
juju deploy mediawiki mywiki juju add-relation mywiki:db masterdb:db juju add-relation mywiki:slave slavedb:db
Upgrade Groups and/or Config Groups
There are also interesting use-cases for breaking large applications down into separate groups of units. Instead of a single 5000-node hadoop application named hadoop-slave, you might build that cluster from multiple smaller application groups.
juju deploy hadoop hadoop-master juju deploy hadoop hadoop-slave-A -n2500 juju deploy hadoop hadoop-slave-B -n2500 juju add-relation hadoop-master:namenode hadoop-slave-A:datanode juju add-relation hadoop-master:namenode hadoop-slave-B:datanode ...
These application groups can be managed independently by Juju for upgrades and configuration
juju config hadoop-slave-B some_param=new_value
This technique can potentially be a way for Juju to manage rolling upgrades for an application. Of course, this depends heavily on the applications in question and how well they support version management, schema changes, etc.