Flaky: watching a real share sometimes never subscribes (stream stuck on Buffering) #573

Open
opened 2026-11-14 21:15:18 +00:00 by charles · 0 comments
Owner

What happened. In CI run 300 (Linux job 7749, branch ci/27-linux-cache, commit db34152) the LiveKit interop test watching_a_real_share_puts_its_frames_on_the_bus (apps/fluxer/tests/voice_livekit.rs:597) failed: no frame of ada's share reached the frame bus within the 20 s step, and the app's remote stream was still state: Buffering. The same test passed on the same commit in the runs just before (7745) and just after (7752), so this is a race.

Evidence from the log (draft release ci-log-300-linux):

  • 20:37:27Z — ada's sender is encoding normally (encode_fps: 27, target_bps: 2500000).
  • 20:37:36Z — the SFU sends ada a subscribed-quality update for her track: vp8:[Low=false, Medium=false, High=false]; dynacast logs layers changed -> [(High)=off].
  • From then on encode_fps: 0, target_bps: 0 until the test times out at 20:37:56Z.

So the server believed nobody was subscribed to the share: the watcher's subscription never took effect, or was dropped right after.

To find out.

  • Whether the race is in the test rig or in the app's "start watching" path (subscribe request sent before the track publication is known, or an unsubscribe/resubscribe pair arriving out of order).
  • If it is the app, a user would see a stream tile stuck on buffering until reopening it — check against #237-style reports.

To do. Reproduce by looping the test (FLUXER_LIVEKIT_TESTS=1 cargo test -p fluxer --test voice_livekit -- --ignored --test-threads=1 watching_a_real_share), log the subscription requests and the SFU's replies, fix the ordering, and keep the test strict (no retry that hides it).

**What happened.** In CI run 300 (Linux job 7749, branch `ci/27-linux-cache`, commit db34152) the LiveKit interop test `watching_a_real_share_puts_its_frames_on_the_bus` (`apps/fluxer/tests/voice_livekit.rs:597`) failed: no frame of ada's share reached the frame bus within the 20 s step, and the app's remote stream was still `state: Buffering`. The same test passed on the same commit in the runs just before (7745) and just after (7752), so this is a race. **Evidence from the log** (draft release `ci-log-300-linux`): - 20:37:27Z — ada's sender is encoding normally (`encode_fps: 27`, `target_bps: 2500000`). - 20:37:36Z — the SFU sends ada a subscribed-quality update for her track: `vp8:[Low=false, Medium=false, High=false]`; dynacast logs `layers changed -> [(High)=off]`. - From then on `encode_fps: 0`, `target_bps: 0` until the test times out at 20:37:56Z. So the server believed nobody was subscribed to the share: the watcher's subscription never took effect, or was dropped right after. **To find out.** - Whether the race is in the test rig or in the app's "start watching" path (subscribe request sent before the track publication is known, or an unsubscribe/resubscribe pair arriving out of order). - If it is the app, a user would see a stream tile stuck on buffering until reopening it — check against #237-style reports. **To do.** Reproduce by looping the test (`FLUXER_LIVEKIT_TESTS=1 cargo test -p fluxer --test voice_livekit -- --ignored --test-threads=1 watching_a_real_share`), log the subscription requests and the SFU's replies, fix the ordering, and keep the test strict (no retry that hides it).
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
charles/fluxer-native#573
No description provided.