Stabilize hosted live video playback
Some checks are pending
ci-gates / checks (push) Waiting to run
deploy-cloudflare / checks (push) Waiting to run
deploy-cloudflare / deploy (push) Blocked by required conditions

This commit is contained in:
every.channel 2026-05-03 22:38:39 -07:00
parent 6739b424ab
commit bd5d9857ed
No known key found for this signature in database
4 changed files with 83 additions and 67 deletions

View file

@ -8,12 +8,15 @@ Hosted live playback can subscribe to a local HDHomeRun stream while the visible
## Decision
Publish WebTransport fMP4 with keyframe-duration fragments (`frag_keyframe`) instead of one fragment per frame, and default `wt-publish` / `nbc-wt-publish` to the non-passthrough CMAF sample path. On the hosted web player, remove the `muted` attribute and reapply unmuted state to both the wrapper and underlying video element after a user gesture.
Publish WebTransport fMP4 with keyframe fragments where every emitted video frame is a keyframe (`g=1`, `keyint_min=1`). For HDHomeRun-style live input, cap the default WebTransport transcode to 6 fps so the hosted watcher receives independently decodable video groups at a sustainable cadence. Expose publisher knobs for the ffmpeg video filter, GOP interval, and fMP4 movflags so runtime operators can tune without another code edit. Keep `wt-publish` / `nbc-wt-publish` on the non-passthrough CMAF sample path. On the hosted web player, render live video through the `<moq-watch>` canvas/WebCodecs path instead of the `<video>`/MSE path, remove the `muted` attribute, and reapply unmuted state to the watcher backend after a user gesture.
## Consequences
- Browser MSE receives continuous segment-duration ranges instead of isolated zero-length frame ranges.
- Live playback accepts GOP-sized fragment latency, matching the current 48-frame live GOP.
- Browser playback receives continuous media ranges without turning each GOP into a single playback jump.
- Live playback and observation diffing receive independently decodable video groups at frame cadence.
- WebTransport video publishing uses more bandwidth per frame, but the 6 fps cap keeps group churn lower than full-rate all-intra OTA publishing.
- Operators can raise or lower `--video-filter`, `--gop-frames`, and `--movflags` from publisher configuration instead of rebuilding.
- Hosted live rendering avoids the upstream MSE path that side-browser validation showed repeatedly skipping slow groups.
- Relay subscribers receive `video0.m4s` and `audio0.m4s` media groups by default instead of catalog-only passthrough announcements.
- Audio remains gesture-gated for autoplay policy, but the gesture now actually unmutes the player.
@ -21,7 +24,11 @@ Publish WebTransport fMP4 with keyframe-duration fragments (`frag_keyframe`) ins
- Raise web jitter again. Rejected because the buffered media ranges were effectively zero-length; more latency does not turn still ranges into playable media.
- Keep passthrough mode as the default. Rejected because relay probes received only `catalog.json` while the non-passthrough sample path delivered video and audio media groups.
- Use keyframe-duration fragments. Rejected after side-browser validation showed repeated `seeking forward/backward` corrections and GOP-sized visual jumps even though audio was healthy.
- Keep 48-frame GOPs with every-frame fMP4 fragments. Rejected because relay archive proof still showed 48-frame video groups; `moq-mux` groups video by keyframe in the non-passthrough path.
- Use full-rate or 12 fps all-intra every-frame fragments. Rejected because relay proof showed one-frame video groups, but side-browser validation produced heavy `skipping slow group` churn from too many tiny media groups.
- Keep the `<video>` child renderer. Rejected because the hosted side browser showed the MSE renderer repeating frames and emitting slow-group skips while subscribed.
## Rollout / teardown
Rebuild/restart local and hosted publishers, deploy the updated web asset, and verify hosted playback by checking frame hashes over time. Teardown is restoring `WT_PUBLISH_MOVFLAGS` to `frag_every_frame`, restoring passthrough defaults to true, and restoring the prior muted wrapper behavior.
Rebuild/restart local and hosted publishers, deploy the updated web asset, and verify hosted playback by checking canvas frame hashes over time plus side-browser console seek corrections. Teardown is setting `--gop-frames 48`, setting `--video-filter` back to the prior source cadence, restoring passthrough defaults to true, and restoring the prior `<video>` child plus muted wrapper behavior.