Hi everyone,
(I CCed -devel and deity, but we probably should just discuss
that on -dpkg)
while breakfast here at DebConf, the topic of delta upgrades
came up. I think delta debs are generally a thing we should
aim to have, but debdelta is not the implementation we want:
* It is not integrated into APT or dpkg
* It relies on a script shipped in the debdelta to generate
a new deb from and old deb
We guessed that generating the new deb from the old deb is
actually the slowest part of the whole debdelta process. I
propose a new solution, which does not really have a name
yet.
# Scope
Patching installed packages to match new state
# Not scope
Generating new package from old package
# Design
Introduce a new format 'ddeb' (delta/differential deb)[1], the
format shall contain a control.tar member and a version member
as in a .deb file. Instead of a data.tar member, it contains
a diff.tar member, however.
The .diff.tar member contains patches to apply to individual
files of the old package. No idea about specific algorithm
to choose here, yet.
The control.tar's control file is extended with an Old-Version
field, so we get some sanity checking. Alternatively, it may
be extended with an mtree of the source we are patching.
The delta files are stored alongside their debs, and referenced
in the packages files in a Patches-<hash> section:
Patches-SHA256:
<hash of delta> <size of delta> <old version> <path>
APT can then grab the delta, if it fails with a 404, it would
fall back to the full delta.
The deltas are not incremental, my suggestion is to do the following ones:
unstable: (1) against testing (2) against previous unstable
experimental: (1) against unstable (2) against previous experimental
stable: (1) against last point release (2) against previous security update
(or rather current one)
All these files will always be around when dak runs anyway, so we do not
need to keep additional historical packages around.
We probably want to make this opt-in, so packages set a field like
Generate-Patches: yes (there might be problems with applying patches to
live systems and bad maintainer scripts).
[1] name should probably change
# Requisites / Extensions to .deb and dpkg database
We need to keep data about the file tree in packages and
in the database, from what I gathered, there already is
a plan to use mtree for that, so that would fit us well.
# Applying a delta deb
Maintainer scripts are run as normally from the
control tarball. The unpack phase is different: Instead
of unpacking a file <file> from the tarball into <file>.dpkg-new,
we apply archive's <file> as a patch to the installed <file>
and store the result in .dpkg-new.
Files not listed in the new mtree but listed in the old
one will be deleted.
# Issues
We need a way to check if we can apply the diff.tar member; and
if we can't, have to download the full deb in APT. This might need
some kind of new patch-prepare state where we basically generate the
.dpkg-new files, but don't apply them (or run any maintscripts).
# Further work
I guess I should do a proof of concept and then we can look if that
is worthwhile, and how it performs.
--
Debian Developer - deb.li/jak | jak-linux.org - free software dev
| Ubuntu Core Developer |
When replying, only quote what is necessary, and write each reply
directly below the part(s) it pertains to ('inline'). Thank you.
