You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
moq-transcode: run() can't bootstrap from a demand-gated source that doesn't advertise its geometry #2788
moq_transcode::run can't bootstrap from a publisher that only encodes while watched, unless that publisher advertises its geometry before encoding anything.
The cycle:
run waits for a catalog snapshot with a usable video rendition, and only subscribes to the source video track afterwards (feed::Feed::new is constructed after the choose_source loop, and even then the subscribe happens lazily in Feed::listen, when a rung is actually demanded).
resolve_rungs needs coded_width/coded_height to size the ladder.
A demand-gated publisher (moq_video::encode::publish_capture, moq-boy) doesn't open its encoder until something subscribes to its video track.
So when the transcoder is the intended consumer, nobody subscribes, no keyframe is produced, the geometry never arrives, and run waits forever. This is the same shape as #2757: a consumer that needs the catalog first, and a publisher that won't produce until demanded.
#2768 fixes the case where the publisher's geometry is known up front (moq capture --width 1920 --height 1080 now advertises it before the camera opens, so the ladder resolves, a viewer subscribes to a rung, and only then does the camera open). The case that remains stuck is a publisher that advertises its codec but not its picture, which is what capture does when you don't pin a size, because a camera's negotiated mode isn't knowable without opening it.
Before #2768 this scenario hung too (capture published no video rendition at all until its first keyframe), so this isn't a regression from that PR. It's the part of the deadlock that publishing the rendition earlier doesn't reach.
Fix direction: give run a bootstrap path that can create demand on the source before it has a ladder, so the source is allowed to produce the keyframe that reveals its geometry. Subscribing to the source video track up front (and dropping it if no rung is ever wanted) would do it, at the cost of opening the camera to learn its mode. Alternatively the ladder could resolve lazily, per rung, once the feed yields its first decoded frame.
#2266 - fix(moq-hls): reconcile catalog renditions [merged] #2439 - fix!: correct catalog, timeline, token, and teardown contracts found in API review [merged] #2608 - feat(bindings): publish raw video with a native encoder [merged] #2630 - fix!: reject a catalog whose broadcast reference escapes the root [merged] #2768 - feat(moq-video): advertise the catalog rendition before the first keyframe [open]
📝 Issue Planner
Check the box below or use the @coderabbitai plan command to generate an implementation plan and prompts that you can use with your favorite coding assistant.
Create Plan
🧪 Issue enrichment is currently in open beta.
You can configure auto-planning by selecting labels in the issue_enrichment configuration.
To disable automatic issue enrichment, add the following to your .coderabbit.yaml:
issue_enrichment:
auto_enrich:
enabled: false
💬 Have feedback or questions? Drop into our discord!
Correction. The claim below that "before #2768 this scenario hung too" is wrong, and so is the framing of this issue as pre-existing.
origin/main's capture_loop opens the camera regardless of demand until its first keyframe, and says so:
// The catalog video rendition only appears once a frame has been encoded (the// importer reads the SPS). Until then we capture regardless of demand so a// catalog-driven subscriber can discover the track and trigger `used()`.// After that we release the camera while unwatched.letmut catalog_ready = false;
That pre-roll is what produces the geometry a transcoder needs, unprompted. So native capture plus transcode does bootstrap on main, and the deadlock described here is introduced by removing the pre-roll, not uncovered by it. Thanks to Codex for catching the bad reasoning.
The underlying ordering problem in run() is still real and still worth fixing (it's what makes the pre-roll load-bearing), but it is a blocker for making capture demand-gated rather than an independent defect. #2768 is being reworked accordingly.
Retiring the quest for this in #3347 and closing with it. #2768 landed in its reworked shape: publish_capture opens the camera once at startup to probe the mode and advertises an exact rendition before any demand, so the rendition choose_source waits on (codec plus dimensions) exists from the first catalog snapshot and run() bootstraps. The "codec but no picture" rendition this issue describes no longer occurs in tree. The remaining resize-between-opens case is #2799.
Activity
coderabbitai commentedon Aug 12, 2026
🔗 Related PRs
#2266 - fix(moq-hls): reconcile catalog renditions [merged]
#2439 - fix!: correct catalog, timeline, token, and teardown contracts found in API review [merged]
#2608 - feat(bindings): publish raw video with a native encoder [merged]
#2630 - fix!: reject a catalog whose broadcast reference escapes the root [merged]
#2768 - feat(moq-video): advertise the catalog rendition before the first keyframe [open]
📝 Issue Planner
Check the box below or use the
@coderabbitai plancommand to generate an implementation plan and prompts that you can use with your favorite coding assistant.🧪 Issue enrichment is currently in open beta.
You can configure auto-planning by selecting labels in the issue_enrichment configuration.
To disable automatic issue enrichment, add the following to your
.coderabbit.yaml:💬 Have feedback or questions? Drop into our discord!
kixelated commentedon Aug 12, 2026
Correction. The claim below that "before #2768 this scenario hung too" is wrong, and so is the framing of this issue as pre-existing.
origin/main'scapture_loopopens the camera regardless of demand until its first keyframe, and says so:That pre-roll is what produces the geometry a transcoder needs, unprompted. So native capture plus transcode does bootstrap on
main, and the deadlock described here is introduced by removing the pre-roll, not uncovered by it. Thanks to Codex for catching the bad reasoning.The underlying ordering problem in
run()is still real and still worth fixing (it's what makes the pre-roll load-bearing), but it is a blocker for making capture demand-gated rather than an independent defect. #2768 is being reworked accordingly.(written by Opus 5)
kixelated commentedon Sep 2, 2026
Retiring the quest for this in #3347 and closing with it. #2768 landed in its reworked shape:
publish_captureopens the camera once at startup to probe the mode and advertises an exact rendition before any demand, so the renditionchoose_sourcewaits on (codec plus dimensions) exists from the first catalog snapshot andrun()bootstraps. The "codec but no picture" rendition this issue describes no longer occurs in tree. The remaining resize-between-opens case is #2799.(written by Claude Fable 5.1)