How anvil is architected

This little HOWTO can be used by those who wish to understand how anvil does things and why some of its architectural decisions were made.


Once upon a time there was a idea of replacing the then existing devstack with a more robust, more error-tolerant and more user/developer friendly OpenStack setup toolkit. Since the existing devstack did (and still does not support very well) complex intercomponent (and interpackage management system) dependencies and installing/packaging/starting/stopping/uninstalling of OpenStack components.

To solve this problem it was thought that there could be a toolkit that could handle this better. It would also be in Python the language of choice for the rest of the OpenStack components thus making it easier to understand for programmers who are already working in OpenStack code. Thus devstack2 was born which was later renamed and after a little while once again got renamed to anvil.


Anvil is designed to have the following set of software components:

  • Actions: an action is a sequence of function calls on a set of implementing classes which follows a logically flow from one step to the next. At the end of each step an action may choose to negate a step of another action.
    • Preparing
      • Downloading source code
      • Post-download patching of the source code
      • Deep dependency & requirement analysis
      • Downloading and packaging of missing python dependencies
      • Packaging downloaded source code into SRPMs (aka source RPMs) that is placed into a SRPM repository.
    • Building
      • Creation of a binary RPM repository with all built packages and dependencies (converting the prepared source RPMs into binary RPMs).
  • Phases: a phase is a step of an action which can be tracked as an individual unit and can be marked as being completed. In the above install action for each component that installed when each step occurs for that component it will be recorded into a file so that if ctrl-c aborts anvil and later the install is restarted anvil can notice that the previous phases have already been completed and those phases can be skipped.
    • This is how anvil does action and step resuming.
  • Components: a component is a class which implements the above steps (which are literally methods on an instance) and is registered with the persona and configuration to be activated. To aid in making it easier to add in new components a set of generic base classes exist that provide common functionality that should work in most simplistic installs. These can be found in anvil/components/. All current components that exist either use these base classes directly or inherit from them and override functions to provide additional capabilities needed to perform the specified action.
  • Distributions: a distribution is a yaml file that is tied to a operating system distribution and provides references for components to use in a generic manner. Some of these references include how to map a components pip-requires file dependencies to distribution specific dependencies (possibly using yum or apt) or what non-specified dependencies are useful in getting the component up and running (such as guestfish for image mounting and manipulation). Other helpful references include allowing for components to specify standard identifiers for configuration such as pip. This allows the underlying yaml file to map the pip command to a distribution-centric command (in RHEL it it’s really named pip-python), see the commands key in the yaml files for examples of these settings. Note that each distribution yaml file that exists in conf/distros provides this set of references for each component and gets selected based on the yaml key in that file named platform_pattern.
  • Configuration: central to how anvil operates is the ability to be largely configuration driven (code when you need it but avoid it if you can). Distributions as seen by the conf/distros folder specify distribution-specific configuration that can be referenced by standard keys by a given component. Each component also receives additional configuration (accessible via a components get_option function) via the yaml files specified in conf/components which provides for a way to have configuration that is not distribution specific but instead is component specific (say for configuring nova to use kvm instead of qemu).
    • This configuration drive approach (as much as can be possible) was a key design goal that drives how anvil was and is developed. It has even seemed to be ahead of its time due to how anvil has a distribution yaml file that has specified component dependencies long before the OpenStack community even recognized such a dependency list was useful.
  • Personas: a persona is a way for anvil to know what components (and possibly subsystems of those components) you wish to have the given action applied to. Since not everyone can agree on what is an install of OpenStack this concept allows for those who wish to have a different set to do so. It is as all other configuration another yaml file and can be examined by looking into the conf/personas folders.
    • Each yaml file contains the list of components to be performed for the given action, a simple set of options for those components (for options that may not be applicable to be in the component configuration yaml) and which subsystems a given component will have enabled (if the component supports this concept) as well as which distribution the persona supports (if there is a desire to restrict a given persona to a given distribution this field can be used to accomplish that goal).