DeveloperWiki:Automated Package Build System
- 1 What is an Automated Build System
- 2 Naming
- 3 Design Proposal
- 4 Programming Language
- 5 Builder Process
- 6 Component Build Out
- 6.1 Master Server Components
- 6.2 Build Server Components
- 6.3 Builder Considerations
- 6.4 Standalone Components
What is an Automated Build System
One of the best examples of an automated package build systems is the Fedora Koji project, Koji is a continuous build system for all of the rpms in the Fedora and RHEL projects. The main benefit of an automated build system is that all of the packages need to pass through a common gate, a common checkpoint for quality and consistency.
While the Koji build server is referenced heavily the Arch equivalent will need to be very different, in both architecture and the general goals of the project.
The proposed name for the Arch package build system is Quarters, the logic being that what really feeds pacman is quarters.
Alternative naming options will be considered
One of the main aspects of the Arch distribution compared to distributions like Fedora is money. The Fedora project can go out and buy servers by the arm full, but Arch will need to continue to rely on volunteer equipment and working with what we can get our hands on. While these aspects will effect the nature of the package build system, they cannot allow for the compromise of the quality of the Arch Linux distribution.
- Avoid requiring too many daemons
- No Authentication systems
- No database dependencies
- Simple https communication
- Able to interact with distributed components over any reasonable link quality
- Need to be able to distribute the build load to systems all over the world
- All communication needs to be encrypted (duh)
- Builders make decisions by peer review
- Data Model
- Information on the available packages is made by parsing live pacman data (pacman is fast enough)
- Packages to build is presented to the builders via serialized format (probably json)
- All build communication is based on gathering presented data, all pulls, not puts
- All servers are state machines, the status of the distributed build environment is assessed by parsing the "global" state machine
- Command system to move packages manually from the build repos to the final repos
- Build Cleanliness
- Every package is built in its own clean chroot environment
- Packages can be pulled and polled from many sources
- AUR (plugable?)
- Other scms ()
- Web interface
- Building Requirements
- All packages need to be build-able, this includes the base toolchain
- Trigger mass rebuilds of package, with version bumps
- Track Packager data through the build process, don't allow the packager to be lost to the builder
- System Interaction points
- Detached cli interface sends signals to the master
- Master server reads signals and acts on them
It has been proposed that the application be developed in Python. Python is fast enough, has the libs we need, and who doesn't understand python?
Sorry, but this project is too big for bash, and bash is too slow.
I thought about OCaml, but I figured that it should be something that everyone can hack on.
Try and go for python3
The process that the builders will use is based on state information, the master server presents the master state, which is the list validated builders and the packages to be built; Only packages which have met all deps and makeedeps in binary repos will be presented. The master server will only present packages which are ready to be built. The builders then download the master server's builder information. The builders pick a package to build, post the package that they have claimed and then query the other builders to see if any other builder has claimed it. If the builder needs to change the package to build then it will just post a claim on a new package. Unless it is the first build on the builder then the peer process of determining which package needs to be built will be done while a package is already being built.
The master will regularly poll the states of the builders, the builders will post their states for the master and for the peers. The builder states will have all the information that the master needs to retrieve the built packages or the information for the failed build.
Component Build Out
Quarters will consist of a number of individual components. These components are all pieces of the distributed system, the components will also interact via a common language independent medium and https.
Master Server Components
The components required for the operation of the master server
The core loop or the core of the master does what most daemons do, takes in the config information, does the double pid form magic, etc.
The core loop will primarily act by continually querying the distributed components for changes, thee components are the builders and the package parsers
The package parsers will be pluggable modules that return predictable data about the state of source package data stored in either a source control manager or an interface such as the abs or the AUR.
The idea here is that as long as the data queried from the source is uniform there can be any number of usable interfaces. The primary system will need to be svn, since this is what the main package tree is stored in, then git and then an AUR parser. More interfaces can be added on later if we or anyone else wants them.
All parsers need to return a similar data structure and may need to do some package preparation. The data structure will need to be defined and static, but it should just be a python dict, something json likes.
This is the main package parser, the parser needs to be able to ready the PKGBUILDS for specific info, particularly pkgver, pkgname and pkgrel, then compare the values to the list of packages that are already built. Once the packages that need updating have been found the parser needs to return a data structure containing the packages that need to be built. The main challenge here is figuring out the fastest way to parse an SVN repo, they can be somewhat slow.
For full functionality most parsers will need companion manipulators, these are interfaces that can manipulate the PKGBUILDS in the underlying SCM. This is needed for mass rebuilds, so that certain packages can trigger deps to auto increment in the SCM. Once the PKGBUILDS have been auto incremented then they will be picked up for rebuilding.
Build Server Components
The build server will run on a host of different nodes. The build server needs to have a number of capabilities. Primarily it will need ample disk space to maintain multiple chroot environments as well as have enough ram and free cpu power.
The Build Server will need to be able to continually get a fresh state from the master.
Peer Review System
The builders will need to be able to download the build state from all of the peer builders. The list of builders will be provided by the master.
After download the builder will need to analyze the state of the other builders and determine what package(s) that this builder should claim. The builder should always have package names ready, so long as packages are being presented by the master.
The builder will need to maintain a single top level loop that controls all of the builder processes and continually queries the master and executes peer reviews.
The main loop will be able to spawn the configured number of chroot builders at the same time. These builders need to be maintained, inasmuch that the number of available packages to build needs to be kept ahead of the builders.
The builders will need to support many simultaneous chroots. A savvy admin for instance could use tmpfs file systems or ssd cards to maintain more and higher speed chroots. So the number of simultaneous package builders needs to be configurable with no limit. Disk IO will be the holding factor for many builds, since fresh chroot copies and updates need to be performed frequently.
When spawning a package builder, the python threading module cannot be used, Quarters as a whole will require multiple processes to operate properly. The threading module, and subsequently the python GIL should be avoided. Instead the multiprocessing module should be used.
While unexpected, the builders need to be protected from malicious code intended to compromise the system. When chroots are constructed the top chroot dir needs to have the barrier attribute test with setattr. Also the builds should be run as an unprivileged user inside the chroot. These precautions will defend against the two main ways to break a chroot, pwd manipulation and dev node hijacking.
These components are used by multiple services
Each component will need to be able to present files via an https server. The package will require a standalone https server written in python.
Quarters will also need to be able to allow for any http server to be used if the admin choses to use something other than the python https server
Nearly all components will need to interact with pacman, he last I checked there we problems with the python bindings for alpm, if I am mistaken, I am all ears.
Some standard functions with respect to pacman will need to be made available.
I know, this is a bottomless pit, the requirements for the PKGBUILD parser are not extensive, although a number of parameters will need to be reliably extracted.
Standard Utils Module
That anoying module that sets up logging and other more globally needed functions