- autobuild-team 0 tasks
- openstack-developers 1 task
- spal-maintainers 0 tasks
Involved Projects and Packages
contains core packages that strongly depend on each other
to provide prjconf and access to old Tumbleweed binaries
on top of ring0 (bootstrap)
See https://en.opensuse.org/openSUSE:Reproducible_openSUSE/Part2 for how to build this with `pbuild`
To test for packaged that embed uname -r in their build results.
This project was created for package pesign-obs-integration via attribute OBS:Maintained
this project should not have packages but only apply extra tweaks via prjconf
to distinguish the effects
from more plain home:bmwiedemann:reproducible
and from home:bmwiedemann:reproducible:test with extra patchy packages
via prjconf
usage:
for r in rb_j1 rb_future1y openSUSE_Factory ; do osc api "/build/home:bmwiedemann:reproducible:rebuild/$r/x86_64/_repository?view=binaryversions" | grep libzzip; done
https://bugzilla.suse.com/show_bug.cgi?id=1237231
https://github.com/openSUSE/obs-build/commit/b801c174d64b32ee7ff8870565990b35c45e4e58
as documented in https://www.starlab.io/blog/adding-sha256-digests-to-rpms
https://en.wikipedia.org/wiki/Year_2038_problem
We try to fix all of it - first in SLFO and then in the rest of openSUSE.
note: vibe-coded, use as inspiration, needs review + testing before real-world use
##Summary
Result: zypper info bash 726 ms → 40 ms (18×, −95%) with stock libzypp, output byte-identical to stock zypper across an 18-case matrix (--xmlout, -q, de_DE locale, exit codes, Ctrl-C).
How it works — reliability first, as requested:
zypperd (new binary linking the existing zypper_lib) loads the pool once using libzypp's readonly hack (never takes the zypp lock, so normal zypper/PackageKit run unaffected), then forks per request: each worker inherits the pool copy-on-write and runs the completely stock Zypper::main() pipeline — no state-reset bugs, worker crashes contained.
zypperc is plain C (libc only, ~1 ms overhead): it passes its argv plus its real stdin/stdout/stderr via SCM_RIGHTS, so isatty/width/colors/prompts are indistinguishable from direct zypper, and Ctrl-C forwards to the worker (verified: graceful stop, rc 105). On any failure before the daemon's "started" frame it silently execs /usr/bin/zypper — a dead daemon costs you stock speed, never a failure.
Never stale data: before each fork the daemon stats rpmdb + repos.d + services.d + solv cookies. rpmdb change → parent Target::reload() (verified live with a package install and removal); repo metadata change → decline + re-exec (0.7 s rewarm of 114k solvables, verified).
Write permissions for root: socket is 0666, but mutating commands need peer uid 0 via SO_PEERCRED — the worker then execs the real zypper (real lock, real prompts on your tty). The netcat bonus works: printf "info bash\n" | nc -U /run/zypperd/zypperd.sock streams output plus a ### exit 0 trailer, read-only commands only.
Bonus find: a genuine libzypp bug — a forked child exiting via exit() (zypper's SIGPIPE path, e.g. zypperc info bash | head) hangs forever in ~LogThread joining the parent-only log thread. Fixed on the libzypp branch (detach when getpid() differs); upstream-worthy since it bites any forking libzypp consumer.
Where everything lives:
zypper repo branch zypperd (off tag 1.14.98): 3 commits (daemon / client / systemd units, socket-activated without a libsystemd dep)
libzypp repo branch zypperd (off 17.38.14): the fork-safety fix
OBS home:bmwiedemann:zypperd rev 2: zypper (Patch1–3, new zypperd subpackage) + libzypp (Patch1) — all green on Tumbleweed i586 + x86_64
zypperd-design.md — design/results doc (left uncommitted in the outer repo for your review)
To try it on a real system: install the zypperd RPM from the project, systemctl enable --now zypperd.socket, then zypperc info bash. Locally the daemon still runs as your user via zypperd --socket /tmp/zd.sock (test harness: ~/.cache/zdtest-diff.sh). The residual 40 ms is child-side option parsing + Augeas zypper.conf + repos.d re-read — cacheable parent-side if you ever want to chase it, and combining with your pool-snapshot patches would cut the re-exec rewarm from ~0.7 s to ~0.2 s.
Test repo setup for building errbit on Leap 16.0 using gems from home:darix:apps
home:gygyegy22w7612:branches:openSUSE:Factory:staging