Fuel Architecture

Fuel ArchitectureΒΆ

Good overview of Fuel architecture is represented on OpenStack wiki. You can find a detailed breakdown of how this works in the Sequence Diagrams.

Master node is the main part of the Fuel project. It contains all the services needed for network provisioning of other managed nodes, installing an operating system, and then deploying OpenStack services to create a cloud environment. Nailgun is the most important service. It is a RESTful application written in Python that contains all the business logic of the system. A user can interact with it either using the Fuel Web interface or by the means of CLI utility. He can create a new environment, edit its settings, assign roles to the discovered nodes, and start the deployment process of the new OpenStack cluster.

Nailgun stores all of its data in a PostgreSQL database. It contains the hardware configuration of all discovered managed nodes, the roles, environment settings, current deployment status and progress of running deployments.


Managed nodes are discovered over PXE using a special bootstrap image and the PXE boot server located on the master node. The bootstrap image runs a special script called Nailgun agent. The agent nailgun-agent.rb collects the server’s hardware information and submits it to Nailgun through the REST API.

The deployment process is started by the user after he has configured a new environment. The Nailgun service creates a JSON data structure with the environment settings, its nodes and their roles and puts this file into the RabbitMQ queue. This message should be received by one of the worker processes who will actually deploy the environment. These processes are called Astute.


The Astute workers are listening to the RabbitMQ queue and receives messages. They use the Astute library which implements all deployment actions. First, it starts the provisioning of the environment’s nodes. Astute uses XML-RPC to set these nodes’ configuration in Cobbler and then reboots the nodes using MCollective agent to let Cobbler install the base operating system. Cobbler is a deployment system that can control DHCP and TFTP services and use them to network boot the managed node and start the OS installer with the user-configured settings.

Astute puts a special message into the RabbitMQ queue that contains the action that should be executed on the managed node. MCollective servers are started on all bootstrapped nodes and they constantly listen for these messages, when they receive a message, they run the required agent action with the given parameters. MCollective agents are just Ruby scripts with a set of procedures. These procedures are actions that the MCollective server can run when asked to.

When the managed node’s OS is installed, Astute can start the deployment of OpenStack services. First, it uploads the node’s configuration to the /etc/astute.yaml file on node using the uploadfile agent. This file contains all the variables and settings that will be needed for the deployment.

Next, Astute uses the puppetsync agent to synchronize Puppet modules and manifests. This agent runs an rsync process that connects to the rsyncd server on the Master node and downloads the latest version of Puppet modules and manifests.


When the modules are synchronized, Astute can run the actual deployment by applying the main Puppet manifest site.pp. MCollective agent runs the Puppet process in the background using the daemonize tool. The command looks like this:

daemonize puppet apply /etc/puppet/manifests/site.pp

Astute periodically polls the agent to check if the deployment has finished and reports the progress to Nailgun through its RabbitMQ queue.

When started, Puppet reads the astute.yaml file content as a fact and then parses it into the $fuel_settings structure used to get all deployment settings.

When the Puppet process exits either successfully or with an error, Astute gets the summary file from the node and reports the results to Nailgun. The user can always monitor both the progress and the results using Fuel Web interface or the CLI tool.

Fuel installs the puppet-pull script. Developers can use it if they need to manually synchronize manifests from the Master node and run the Puppet process on node again.

Astute also does some additional actions, depending on environment configuration, either before the deployment of after successful one:

  • Generates and uploads SSH keys that will be needed during deployment.
  • During network verification phase net_verify.py script.
  • Uploads CirrOS guest image into Glance after the deployment.
  • Updates /etc/hosts file on all nodes when new nodes are deployed.
  • Updates RadosGW map when Ceph nodes are deployed.

Astute also uses MCollective agents when a node or the entire environment is being removed. It erases all boot sectors on the node and reboots it. The node will be network booted with the bootstrap image again, and will be ready to be used in a new environment.