GLEP 9: Gentoo Package Update System

Abstract

This document proposes an official package updating system for Gentoo Linux. The Deltup project has been developed for this purpose. [1]

As packages grow larger the amount of redundant data keeps increasing. Updating existing tarballs by patching is the natural way to handle source updates.

Motivation

This system will reduce mirror loads (potentially mirror size as well) and significantly speed up downloads, making Gentoo much more attractive for users with low-bandwidth connections.

Server Implementation

I propose that the patches be put onto the Gentoo Mirrors and stored in a new directory called "patchfiles" which could be placed beside "distfiles".

It would be advantageous to have a list of available patches within the portage tree so that it can be updated during "emerge sync". A file named "dtu.list" can be created and placed in $PORTDIR/profiles.

If a machine can be set up to generate patches it should contain a local mirror of distfiles which it can monitor for added packages. When a package is added to distfiles the machine can try to determine the previous tarball so a patch can be made and placed in the patchfiles dir. In addition, special-case patches can be added manually.

The dtu.list file will be maintained by a special script. Whenever patches are added or removed to the patchfiles dir, the script will make necessary additions/removals in dtu.list. This will be done with minimal changes in the file so it can be synchronized efficiently.

Client Implementation

The system will be optional for users and can be enabled by making portage invoke efetch through the FETCHCOMMAND environment variable [3].

When a package fetch is requested, the efetch/fetchcommand scripts (part of Deltup) will scan the dtu.list file for updates and try downloading and applying them if they exist, or fall back to a full package download if they don't or if the patching process fails.

Rationale

The most controversial feature has been the addition of dtu.list to the portage tree, so in this section I will list the reasons I support it.

• Flexibility. Without it, there must be a standard naming scheme which we would be stuck with once the system is in place. Changing the system would require serious compatibility breaks. With the dtu.list file we can change the naming scheme easily without problems, or even have several different naming schemes.
• Features. Without patch information detecting different upgrade paths would be impossible. Split package patching would also be impossible. If the info is available we can use it to find the quickest upgrade path, like jumping from a .0 release, or even disable certain patches if there are problems with them.
• It would be impossible to know which packages to upgrade from in some cases, including renamed packages.
• Knowing which patches are available will eliminate the overhead of attempting to download patches which don't exist.

The dtu.list file will contain several hundred kilobytes of data. That has caused some concern over how efficiently it can be rsynced. To address these concerns the file's format will be plaintext and care has been taken to minimize the number of changes as removals/additions are made.

Backwards Compatibility

There are no backwards compatibility issues since Deltup can generate correct package MD5sums.

Conclusion

I suggest we start with a scaled-down implementation and provide more as the demand increases. All of the necessary code is already written and working in non-official tests.

References

Copyright

This work is licensed under the Creative Commons Attribution-ShareAlike 3.0 Unported License. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/3.0/.