Goal is to create a system that we can use to scale the thebe backends. The idea is to:

• Mimic the tmpnb API for spawning an image, except that we can also spawn a specific container
• Use docker swarm as a backend; in this case, we'll also use Rackspace's hosting service
• Use interlock (https://github.com/ehazlett/interlock) to handle service discovery to the Docker containers

The following diagram summarizes the architecture:

All the docker containers are launched and managed by on a swarm cluster managed by Rackspace. We'll use the Rackspace offering for this as a host (https://mycluster.rackspacecloud.com/).

Creating a swarm cluster on Rackspace has the details. If you're unfamiliar with some of the Rackspace tools, the Rack overview will give you a bit of background on the service.

Interlock, a Docker plugin service, provides a slick way to us HAProxy to handle service discovery for clusters on the swarm. So, whenever we start a container on the swarm, Interlock will expose it using the hostname we supply in the docker run command. When you kill a container, interlock will delete it from the host setup.

Service discovery with interlock (HAProxy plugin) has the details.

thebe-server is a Go-based application that provides an API for starting and stopping containers on the swarm. Basically, it provide similar endpoints used by tmpnb, but it does not try to handle pooling or other services. Instead, it's focus is on allowing you to start many different kinds of containers that you can connect to thebe

The thebe-server API has details on the API and on the development and deployment process. The cool things about thebe-server is that it is also a containerized, and can run on the same swarm setup as the other containers, making it simple to manage.