Repository navigation
feat: add Utah's unique @multimedia closure members (PackageKit, alsa-utils, ffmpegthumbnailer, gstreamer1 plugins) - #81
Conversation
Expand the effective Bluefin multimedia transaction into an explicit factory closure, including mandatory @multimedia members, free-codec compatibility Provides, staged source recipes, and a machine-readable NEVRA report. Add validation/tests and durable Hummingbird targeting guidance for source aliases, Hummingbird-owned binaries, and ABI-cycle exceptions. Assisted-by: GPT-5 via GitHub Copilot Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Signed-off-by: Danathar <Danathar@users.noreply.github.com>
hanthor
left a comment
There was a problem hiding this comment.
The work is careful and the gates all pass — but 41 of this PR's 47 package directories are also added by #70, and it raises the runtime contract by 45 packages that have never been built in this factory. Both need resolving before it lands.
Overlap with the other two large PRs. Package directories touched, from three-dot diffs against main (54cc8d1):
#81 packages: 47 #70 packages: 161 #51 packages: 156
#81 ∩ #70 = 41
#81 ∩ #51 = 40
The shared 41 include the entire codec closure this PR is named for:
aribb24, codec2, fdk-aac-free, ffmpeg, flac, gsm,
gstreamer1-plugins-bad-free, highway, jpegxl, lame, libXv,
libaribcaption, libasyncns, libcdio, libcdio-paranoia, libebur128,
libheif, liblc3, libldac, libogg, libsndfile, libtheora,
libvisual, libvorbis, libvpl, ...
So three large PRs are concurrently re-adding substantially the same recipes on three different derivations. They do not textually conflict with main individually, but merging any two would produce two independent source locks, stage assignments and spec variants for the same 40 packages, resolved by merge order. This is the thing to settle first — ideally by deciding whether #70 (which already carries ffmpeg 9.0.1 and has been through a 285-job matrix run) is the base, and rebasing this PR to add only the ~6 packages it uniquely contributes plus the multimedia_overrides contract change.
The contract expansion is large and unproven. Merged into main locally:
PR#81 fc=0 val=0 pytest=0 | validated 238 source RPMs | 78 passed in 0.66s
PR#81 quoting=0 runtime_contract=0 unittest=0 | validated runtime contract: 117 packages
Baseline on unmodified main is 193 SRPMs and 72 contract packages. So this adds 45 recipes and asserts 45 more packages must resolve, via:
-# multimedia_overrides joins this list when utah-packages#24 closes. The
-# baseline transaction first proves Utah's currently enabled contract; adding
-# a knowingly incomplete future contract here would prevent even diagnostic
-# baseline publication without identifying which policy was being tested.
+# The explicit multimedia_overrides expansion is part of Bluefin's effective
+# install transaction, including the mandatory members of @multimedia.
+ "multimedia_overrides",The comment being deleted is an explicit argument against doing exactly this, and it names the condition — "when utah-packages#24 closes". The PR body does not say #24 has closed or why that reasoning no longer applies. Worth addressing directly: if the multimedia closure is incomplete or any of the 45 fails to build, runtime_contract.py --check in rebuild-rpms.yml's prepare job starts failing and blocks publication of everything, not just multimedia. That is precisely the failure mode the deleted comment was guarding against.
That risk is not hypothetical here. gstreamer1-plugins-bad-free is one of the 41 shared packages and is a confirmed failure on unmodified main (run 34571911239 @ 54cc8d1, job 103185436349). #70 addresses it with a dedicated chromaprint bcond, reasoning that Fedora's libchromaprint links libavutil.so.60 while this factory's ffmpeg provides so.61. I do not see an equivalent change here, so merging this PR's contract expansion with that recipe still failing would wedge the publish gate.
No CI has run on this. get_status returns total_count: 0, state: pending — it is a fork PR (Danathar/utah-packages) awaiting workflow approval. The body's validation list (just check, just test 78 passed, runtime_contract --check, validate_multimedia_closure.py, "targeted Fedora 44 + Hummingbird spec preflight") all reproduces locally, and I confirm 78 passed exactly. But none of it builds an RPM. For 45 new recipes in a factory where roughly 15 of 193 existing recipes currently fail to build, a real matrix run is the evidence that matters.
What is good here. The validate_multimedia_closure.py validator plus the machine-readable 45-binary closure report is the right instinct — it makes the claim checkable rather than asserted, and it is the piece most worth keeping regardless of how the overlap is resolved. The versioned compatibility Provides for the unsuffixed Bluefin ffmpeg and fdk-aac-free transaction names is a real detail that the other two PRs do not appear to handle, and documenting the Hummingbird-provided alsa-ucm and the conditional eSpeak exception is genuinely useful. If this PR is rebased down to its unique contribution, those four things are what it should carry.
What I would ask for:
- Settle the three-way overlap with #70 and #51 before merging any of them; rebase this onto whichever wins and reduce it to its unique packages.
- Either land the
multimedia_overridescontract line in a separate follow-up after the recipes are proven to build, or explain why the deleted comment's reasoning no longer holds. - Get one real matrix run on the 45 new recipes.
Generated by Claude Code
hanthor's review found 41 of this PR's 47 new package directories also appear in projectbluefin#70 and projectbluefin#51 (both further along). Rebase down to the 6 packages unique to this PR (PackageKit, alsa-utils, ffmpegthumbnailer, gstreamer1-plugin-libav, gstreamer1-plugins-ugly-free, rust-gst-plugin-dav1d), and: - Revert the gstreamer1-plugins-bad-free and libheif spec edits, since projectbluefin#70 addresses the same gstreamer1-plugins-bad-free build failure with a different fix (a chromaprint bcond). - Defer the [bluefin].sections multimedia_overrides contract-enforcement change in config/runtime-contract.toml, restoring the original comment explaining why — the 39 dropped recipes are unproven here. - Trim config/bluefin-packages.toml's [multimedia_overrides] list, .packit.yaml, and config/upstream-sources.json to only the 6 kept packages (plus the pre-existing Mesa/Intel/VA-API overrides). - Trim reports/bluefin-multimedia-closure.json to the 19 requirements this PR actually builds, and reframe docs/skills/multimedia-closure.md and docs/targeting-hummingbird.md as documenting the target 45-binary closure, with the remainder expected via projectbluefin#70. tools/validate_multimedia_closure.py, tools/import_bluefin_rawhide.py, and their tests are kept as-is — generic infrastructure, not tied to package count. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01LNgJ8VMdoqzGequbxsP2Kc
|
@hanthor Thanks for catching the overlap — confirmed it independently (three-dot diffs against Pushed
One thing I couldn't carry forward and want to flag explicitly rather than let it quietly disappear: the original diff added versioned compatibility
|
hanthor
left a comment
There was a problem hiding this comment.
Re-reviewed 3b692c0 — every claim verified, approving
I checked each item against the tree rather than the description.
Scope is genuinely cut to 6 packages. git diff --name-status origin/main... shows exactly the six you list and nothing else under packages/:
A packages/PackageKit
A packages/alsa-utils
A packages/ffmpegthumbnailer
A packages/gstreamer1-plugin-libav
A packages/gstreamer1-plugins-ugly-free
A packages/rust-gst-plugin-dav1d
And they really are unique. I ran each of the six against #70 and #51 individually — zero changed files in either PR for all six. The overlap is gone, not just reduced.
The reverts are real, not partial. git diff origin/main... for config/runtime-contract.toml is now empty output — the multimedia_overrides contract-enforcement change and its comment are fully restored, not edited back to something approximate. gstreamer1-plugins-bad-free and libheif no longer appear in the diff at all, so there's no conflicting chromaprint fix racing #70.
The trimmed closure report is internally honest and actually gated. tools/validate_multimedia_closure.py reports 19 built, 1 Hummingbird-provided, matching your count. I confirmed it isn't decorative:
- drop one entry from
requirements→ERROR: report requirements do not match multimedia_overrides.packages (missing=['mesa-vulkan-drivers'], extra=[]), exit 1 - clean tree → exit 0
tests/test_multimedia_closure.py::test_committed_report_matches_factory_contractgoes red on the same tamper
and .github/workflows/validate.yml runs python3 -m pytest tests -q, which imports validate through that test — so the gate is live in CI, not just a script someone has to remember to run.
python3 -m pytest tests/ -q— 78 passed
The Provides: detail you flagged
This is the part worth calling out: surfacing the dropped versioned compatibility Provides: for the unsuffixed ffmpeg/ffmpeg-libs and libfdk-aac transaction names, instead of letting it vanish with the reverted specs, is exactly right. A scope reduction is the easiest place in the world to silently lose a one-line fix that nobody notices until a transaction fails to resolve.
I've noted it on #70 so it gets checked before that one lands rather than after.
CI
Still action_required — fork PR awaiting a maintainer-approved run, same as the other ten in this repo. Not a property of this diff. Local gates pass: just check (199 source RPMs), just test (78), and the runtime contract check all clean on this head.
Generated by Claude Code
|
Closing as stale rather than waiting on a rebase: it conflicts with Generated by Claude Code |
Summary
@multimediamembers unique to this PR after deduplicating against the two other large codec/closure PRs (Build the whole Utah contract: 343 recipes, Firefox as a Flatpak, four stage failures fixed #70, Add the 152 recipes a Utah image needs, on the Packit inventory shape #51), which both add substantially the same recipes:PackageKit-gstreamer-plugin,alsa-utils,ffmpegthumbnailer,gstreamer1-plugin-dav1d(viarust-gst-plugin-dav1d),gstreamer1-plugin-libav, andgstreamer1-plugins-ugly-free.reports/bluefin-multimedia-closure.json), a validator (tools/validate_multimedia_closure.py), and a procedure doc (docs/skills/multimedia-closure.md) covering the members this PR actually builds, with the rest tracked as expected via Build the whole Utah contract: 343 recipes, Firefox as a Flatpak, four stage failures fixed #70.alsa-ucmand the conditionalgstreamer-plugins-espeakexception.tools/import_bluefin_rawhide.py's generic source-override support (used here for the pre-existinglibva-intel-media-driver→intel-media-driver-freemapping).What changed from the original version of this PR
hanthor's review (#81 (review)) found that 41 of this PR's original 47 new package directories also appear in #70 and #51, and that the
multimedia_overridescontract-enforcement change (config/runtime-contract.toml) would assert a 45-binary closure that isn't proven to build —gstreamer1-plugins-bad-freein particular is a confirmed build failure on unmodifiedmain.In response:
gstreamer1-plugins-bad-freeandlibheifspec edits — Build the whole Utah contract: 343 recipes, Firefox as a Flatpak, four stage failures fixed #70 fixes the samegstreamer1-plugins-bad-freebuild failure with a different approach (a dedicated chromaprint bcond).config/runtime-contract.tomlchange that addedmultimedia_overridesto the enforced contract sections, restoring the original comment explaining why that should wait until the closure is proven.config/bluefin-packages.toml's[multimedia_overrides]list,.packit.yaml, andconfig/upstream-sources.jsonto the 6 kept packages (plus the pre-existing Mesa/Intel/VA-API overrides, untouched).Not carried forward, flagged for #70 instead: the original diff added versioned compatibility
Provides:lines on the Bluefin-unsuffixedffmpeg/ffmpeg-libs/libfdk-aactransaction names inpackages/ffmpeg/ffmpeg.specandpackages/fdk-aac-free/fdk-aac-free.spec. Since those specs are dropped here (duplicated by #70), that detail needs to land in whichever PR ends up carrying those recipes.Validation
just check— passes (199 source RPMs, factory contract satisfied)just test— 78 passedpython3 tools/runtime_contract.py config/bluefin-packages.toml config/runtime-contract.toml --check— 72 packages (unchanged frommain, confirming the contract-enforcement change is fully reverted)python3 tools/validate_multimedia_closure.py— 19 built, 1 Hummingbird-provided