Repository navigation
Conversation
Today's scheduled rebuild failed 7 of 8 packages (samba, webkitgtk, evolution-data-server, gjs, firefox, pipewire-libs-extra, gstreamer1-plugins-bad-free) with the same conflict: Problem 1: cannot install both libicu-78.3-1.hum1.bfin.x86_64 from stages and libicu-77.1-2.1.hum1.x86_64 from public-hummingbird-x86_64-rpms icu is hardcoded into HB_EXCLUDE (to always prefer Hummingbird's icu over Fedora's), but the factory now also builds icu itself at stage 0 (config/upstream-sources.json, version 78.3 — newer than both Fedora's and Hummingbird's). The EXCLUDE variable built from actually-staged RPM names (/work/prior) was only ever applied to fedora.excludepkgs and updates.excludepkgs, never to public-hummingbird-x86_64-rpms.excludepkgs — so once a package name exists in both the local stages repo and the Hummingbird repo, dnf sees two unreconciled providers. Fix: split the combined $EXCLUDE into $STAGE_EXCLUDE (names actually built in prior stages, from rpm -qp on /work/prior) and the merged $EXCLUDE ($HB_EXCLUDE + $STAGE_EXCLUDE) used for fedora/updates as before. $STAGE_EXCLUDE alone is now also applied to public-hummingbird-x86_64-rpms.excludepkgs on every builddep call, in every stage. This keeps the hardcoded HB_EXCLUDE names (ruby gems, gpgme, qt6-qtbase, and icu/libicu when NOT locally staged) sourced from Hummingbird as intended, while ensuring anything the factory actually builds always wins over Hummingbird's own copy of the same name, not just over Fedora's. Verified failure: https://github.com/projectbluefin/utah-packages/actions/runs/33950836585 (all 7 failures share the identical libicu conflict; the 8th failure, mozjs140, is an unrelated rust-target-triplet issue not addressed here) Assisted-by: Claude Sonnet 5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
08b3811 to
4ccd301
Compare
Signed-off-by: sec-check[bot] <sec-check[bot]@users.noreply.github.com>
hanthor
left a comment
There was a problem hiding this comment.
The diagnosis is sound and the PR is genuinely small — unlike several others in this queue, the diff matches the title. It just needs a rebase: main has moved and the branch no longer merges.
Merge state. GitHub reports mergeable_state: dirty, and locally:
$ git merge pr/39
CONFLICT (content): Merge conflict in .github/workflows/rebuild-rpms.yml
Base is 9dc62b5; main is now 54cc8d1. Three-dot diff is a reasonable 3 files, +103/−25, so this is an ordinary rebase rather than a drifted branch.
Because it does not merge, I could not run the repo's gates against it. Baseline on unmodified main for comparison when you do rebase: factory_contract rc=0, validate rc=0 (193 SRPMs), pytest rc=0 (72 passed), check_workflow_quoting rc=0 (10 workflows), runtime_contract rc=0 (72 packages).
The root cause holds up. The $EXCLUDE list built from staged RPM names is applied to fedora.excludepkgs and updates.excludepkgs and never to public-hummingbird-x86_64-rpms.excludepkgs — I confirmed that on current main; both builddep call sites (lines 207 and 230 of build-stage.yml) still have exactly those two setopts. So the libicu 78.3-from-stages vs 77.1-from-Hummingbird collision you describe is still live. Splitting $STAGE_EXCLUDE from $HB_EXCLUDE and applying the former to the Hummingbird repo is the right shape: hardcoded Hummingbird-sourced names keep winning over Fedora, while anything the factory actually builds wins over Hummingbird too.
Worth knowing before you rebase — three other open PRs now edit the same lines.
- #83 adds
--setopt=public-hummingbird-x86_64-rpms.excludepkgs="$HB_REPO_EXCLUDE"to bothbuilddepcalls, for the rubydefault-gemsconflict. Same setopt, same two call sites as your change. - #58 attacks the same ruby problem with a global
--exclude=instead, and additionally addssqlite,sqlite-libstoHB_EXCLUDE. #83 and #58 already conflict with each other (2 hunks). - #70 rewrites
FACTORY_REPOhandling in the same block.
Whoever rebases last inherits all of it. It may be worth coordinating a single PR that settles the exclusion mechanism — per-repo excludepkgs vs global --exclude vs your $STAGE_EXCLUDE split — since all three are solving variants of "two repos offer the same name and dnf picks wrong".
A note on the failures this cites. The run you reference (33950836585) is from 5 September. On the most recent scheduled run of unmodified main (34571911239 @ 54cc8d1) the failure set is:
gcc, dracut, evolution-ews, fish, tailscale, git, openssh,
pipewire-libs-extra, samba, evolution-data-server,
gstreamer1-plugins-bad-free, webkitgtk, mozjs140, gjs, firefox
Six of your seven are still failing (samba, webkitgtk, evolution-data-server, gjs, pipewire-libs-extra, gstreamer1-plugins-bad-free); firefox still fails too. But the causes have partly shifted — dracut and git now die on the Hummingbird ruby default-gems file conflict (job 103175696635: file /usr/bin/erb conflicts between attempted installs of ruby4.0-default-gems-4.0.6-37.3.hum1 and ruby3.3-default-gems-3.3.10-23.6.hum1), not on icu. So it would be worth re-confirming after rebase that the icu collision is still what is blocking your seven, rather than assuming the September diagnosis still applies unchanged.
Your note that mozjs140 is a separate rust-target-triplet issue is correct and still true; it is fixed by the sccache/clang hunk currently duplicated across #56, #57 and #58.
Generated by Claude Code
|
Closing as superseded: Generated by Claude Code |
Fixes 7/8 failures in today's scheduled rebuild: https://github.com/projectbluefin/utah-packages/actions/runs/33950836585
All 7 (
samba,webkitgtk,evolution-data-server,gjs,firefox,pipewire-libs-extra,gstreamer1-plugins-bad-free) failed with the identicalconflict:
Root cause
icu/libicuis hardcoded intoHB_EXCLUDEso Hummingbird's copy always winsover Fedora's. But the factory now also builds
icuitself at stage 0(
config/upstream-sources.json, version 78.3 — newer than both Fedora's andHummingbird's). The
$EXCLUDElist built from actually-staged RPM names(
/work/prior) was only ever applied tofedora.excludepkgsandupdates.excludepkgs— never topublic-hummingbird-x86_64-rpms.excludepkgs.So once a name exists in both the local
stagesrepo and the Hummingbird repo,dnf sees two unreconciled providers and fails.
Fix
Split the combined
$EXCLUDEinto$STAGE_EXCLUDE(names actually built inprior stages) and the merged
$EXCLUDE($HB_EXCLUDE+$STAGE_EXCLUDE) usedfor fedora/updates as before.
$STAGE_EXCLUDEalone is now also applied topublic-hummingbird-x86_64-rpms.excludepkgson every builddep call, in everystage. This keeps the hardcoded
HB_EXCLUDEnames sourced from Hummingbird asintended (ruby gems, gpgme, qt6-qtbase, and icu/libicu when not locally
staged), while ensuring anything the factory actually builds always wins over
Hummingbird's own copy, not just over Fedora's.
Not addressed here: the 8th failure,
mozjs140, is an unrelated rust-target-triplet issue (
Don't know how to translate x86_64-redhat-linux-gnu for rustc) — separate bug, separate fix.Verified: YAML parses, and the modified bash block passes
bash -n. Realverification will come from this PR's own CI run against the affected
packages.