Skip to content

noextract-reverted.md

Latest commit

 

History

History
205 lines (156 loc) · 6.94 KB

noextract-reverted.md

File metadata and controls

205 lines (156 loc) · 6.94 KB
Raw

NoExtract, and why there is none

⛔ A downstream consumer broke. For one build, this image stripped documentation, info pages, gtk-doc and every locale through NoExtract. Every rule is now gone, man pages included, so the package database and the filesystem agree.

The report

Samueru-sama/Anylinux-AppImages, one of the named consumers, failed with:

/usr/share/doc/qt6/global/template/images/Qt-logo.png is NOT a valid path!

The path is in the qt6-base package. Their tooling walks a package's file list and checks each path exists.

Reproduced

podman run --rm --platform linux/amd64 docker.io/pkgforge/archlinux:latest bash -c '
  pacman -Sy --noconfirm --needed qt6-base >/dev/null 2>&1
  pacman -Qlq qt6-base | grep -c "^/usr/share/doc/"
  ls -l /usr/share/doc/qt6/global/template/images/Qt-logo.png'
167
ls: cannot access '/usr/share/doc/qt6/global/template/images/Qt-logo.png': No such file or directory

⭐ The database and the filesystem disagreed. pacman -Qlq listed 167 paths under /usr/share/doc/ for that one package, and none of them were on disk. NoExtract governs what is written; it does not change what the package database records.

The scale, which the symptom understated

Same image, base plus qt6-base, every path the database claims, checked:

pacman -Qlq | while read -r p; do [ -e "$p" ] || printf '%s\n' "$p"; done
paths
database lists 49035
missing on disk 25513

By prefix:

prefix missing
usr/share/man 18944
usr/share/locale 4539
usr/share/doc 1022
usr/share/i18n 563
usr/share/X11 236
usr/share/info 135
usr/share/gtk-doc 68

⚠ 52% of every path pacman claimed was absent. Fixing only usr/share/doc would have left 24491 of them.

What was a regression, measured against the image that preceded the rebuild

ghcr.io/pkgforge-dev/archlinux:v2026.05.15, the newest tag published before the rebuild:

podman run --rm --platform linux/amd64 ghcr.io/pkgforge-dev/archlinux:v2026.05.15 bash -c '
  for d in /usr/share/doc /usr/share/man /usr/share/info /usr/share/gtk-doc \
           /usr/share/help /usr/share/locale /usr/share/i18n /usr/share/X11/locale; do
    [ -d "$d" ] && echo "$d present $(find "$d" -type f | wc -l) $(du -sh "$d" | cut -f1)" || echo "$d ABSENT"
  done'
path before the rebuild after it verdict
usr/share/doc present, 545 files, 13M absent ⛔ regression
usr/share/info present, 104 files, 7.3M absent ⛔ regression
usr/share/gtk-doc present, 43 files, 284K absent ⛔ regression
usr/share/locale present, 1494 files, 109M stripped ⛔ regression
usr/share/i18n present, 604 files, 17M stripped ⛔ regression
usr/share/man ABSENT absent unchanged
usr/share/help ABSENT absent unchanged
usr/share/X11/locale ABSENT absent unchanged

⭐ Man pages were already gone. The image before the rebuild deleted /usr/share/man with a PostTransaction hook, rootfs/any/usr/share/libalpm/hooks/man-page-remove.hook at commit 9d1e142, which re-ran on the consumer's next transaction. Everything else on that list is a change this project introduced.

What is in the tree now

No NoExtract rule at all, in any rootfs/<arch>/etc/pacman.conf.

grep -n '^NoExtract' rootfs/*/etc/pacman.conf

Exit status 1. Only the stock commented #NoExtract = default remains.

⭐ The first fix kept usr/share/man/*, on the grounds that man pages had been absent for years and so were not a regression. The maintainer ruled that the database and the filesystem should agree completely, which meant that rule going too.

⚠ One more removal was found while checking the result. rootfs/any/etc/pacman.d/hooks/package-cleanup.hook ran rm -rf /var/cache/pacman/pkg, which deletes a directory the pacman package owns. It now runs find /var/cache/pacman/pkg -mindepth 1 -delete, which empties it and leaves the directory. Dropping the cached packages is still a sane default: they come back on demand.

Verified

Built from this tree and checked against the report:

podman run --rm --platform linux/amd64 localhost/archlinux:final bash -c '
  pacman -Sy --noconfirm --needed qt6-base >/dev/null 2>&1
  ls -l /usr/share/doc/qt6/global/template/images/Qt-logo.png
  pacman -Qlq | while read -r p; do [ -e "$p" ] || printf %s\n "$p"; done'
-rw-r--r-- 1 root root 2425 Aug 20 07:47 /usr/share/doc/qt6/global/template/images/Qt-logo.png
/var/lock
published, stripped first fix, man only now
database lists 49035 49035 49035
missing on disk 25513 18946 1

⭐ The one remaining is /var/lock, a symlink to ../run/lock owned by the filesystem package. /run/lock is created by systemd, so in a container the link dangles. docker.io/library/archlinux:latest behaves the same way. It is not something this image withholds.

A locale needs no preparation:

echo "de_DE.UTF-8 UTF-8" >> /etc/locale.gen && locale-gen
locale -a
C C.utf8 de_DE.utf8 en_US.utf8 POSIX

tests/run.sh image passes 30 of 30 against the rebuilt image, and tests/static/80-docs-claims.sh assertion 9 now fails on any NoExtract line in any shipped config.

What this costs

podman image inspect <ref> --format '{{.Size}}'
image size
ghcr.io/pkgforge-dev/archlinux:v2026.05.15, before the rebuild 521 MiB
docker.io/pkgforge/archlinux:latest, everything stripped 395 MiB
this tree, nothing stripped 543 MiB

⚠ 543 MiB is 22 MiB above what consumers had before the rebuild, because man pages are now extracted too, and 148 MiB above the stripped build.

⭐ The bar was never size. It is whether a consumer could undo the change, and the answer turned out to be no in the way that mattered: they could restore the files, but not before their build had already failed on a path the package database said was there.

The same complaint exists upstream, and still applies to their image

archlinux/archlinux-docker issue 55, Suggestion: restore excluded files in base image and provide a slim variant, opened 2020-12-06 and still open. The reporter quotes that project's own stated principle back at it:

All installed packages have to be kept unmodified

The official image still behaves that way:

podman run --rm --platform linux/amd64 docker.io/library/archlinux:latest bash -c \
  'pacman -Qlq | while read -r p; do [ -e "$p" ] || printf %s\n "$p"; done | head -8'
/usr/share/doc/
/usr/share/doc/acl/
/usr/share/doc/acl/CHANGES
/usr/share/doc/acl/COPYING
/usr/share/doc/acl/COPYING.LGPL
/usr/share/doc/acl/extensions.txt
/usr/share/doc/acl/libacl.txt
/usr/share/locale/de/

Recorded in HISTORY/references/archlinux-docker.md. It was read before this incident and treated as cost evidence about somebody else's decision.