Skip to content

A consumer for quic-deadline: MPEG-TS trunks under a stated delay #8559

Open

Description

@t0ms

Two m5 quests that #8423 pruned for want of a consumer have one here, should either be re-planned: quic-deadline and, more weakly, quic-careful-resume.

The consumer. A broadcast trunk: a whole MPEG-TS multiplex at a constant 10–25 Mb/s, published by moq import ts (demultiplexed or --passthrough) and received by moq export ts --delay D, which plays out a fixed D behind and already sends D as the subscription's Delay::max (rs/moq-cli/src/subscribe.rs). The incumbent, SRT, delivers whatever can still arrive within its latency and drops what cannot. That is quic-deadline's Goal, and the half a trunk needs most is "below the deadline, recovery gets faster".

A 5 s outage costs a trunk about the outage, whatever the delay. One path at 20 Mb/s (cake) and 100 ms RTT, a 9.95 Mb/s multiplex, moq-relay on BBRv3, every packet in the relay's direction dropped for 5 s, three replicates per cell, graded on the video delivered. SRT is on the same shape at the nearest delivered latency we measured it at (2.18, 5.91 and 11.87 s; libsrt, buffers sized to it):

Build, relay --delay (delivered) Video lost SRT
151fa4845, moq-relay 2 s (2.4 s) 7.68–8.48 s 3.74 s
d66c892833, moq-relay 5.75 s (5.91 s) 5.0–8.2 s 0.18 s
d66c892833, moq-relay 11.71 s (11.87 s) 3.0–6.4 s 0 s

At 11.7 s the budget is more than twice the outage and buys about a second. In all twelve d66c892833 cells the subscriber's log places the loss: its moq-net gave up on the groups the outage had stalled (current group evicted; skipping to next buffered group … Moq(Old)), the first of them 1.9–6.4 s after the link returned, while they still waited on data, and the export's skips account for the video lost. The same client on the same last mile through a moqx relay loses 3.6–5.0 s the same way.

Where the time goes, at 2 s on moq-quic. The relay's next send comes 2.43–2.46 s after the link returns, on its backed-off probe timer. The re-arm we added in #6178 fires during a one-way outage, on the subscriber's packets, and its probe goes into the dead direction. Dropping both directions costs 6.12–12.64 s: the first acknowledgement after the return carries an RTT sample spanning the outage (smoothed RTT 105 ms → 0.74–0.99 s), and the next send comes 0.95–7.76 s after the return. By reading, pto_time_and_space, on_loss_detection_timeout and RttEstimator::update are unchanged on main at ced5a45c5f. Two things a re-plan may want to know:

  • As planned, the early probe never exceeds the congestion window. During an outage the window is full of unacknowledged packets, so it would not fire there, and the outage is this consumer's shape.
  • We have not traced in which order the relay sends the stalled groups and the newer ones after the return; on this shape with the relay's qlog we can.

A re-plan could grade beside bufferbloat on this shape: under a --delay longer than the outage and its drain (11.7 s here), a 5 s outage in either direction costs what SRT's does, with nothing retransmitted past its deadline. We can run the matrix against a branch.

quic-careful-resume, more weakly. The same trunks redial after a relay restart, a GOAWAY handover or a cluster peer's restart, and the receiver joins at a group start up to a GOP old, so it needs the multiplex's rate plus that backlog inside D. We have not isolated the ramp's share: a relay restarted with its reset key kept costs 4.20 s of video (151fa4845, 20 ms RTT, where a ramp is short), most of it the clients learning of the restart (1.7–3.0 s) and the exporter's own restart. At 25 Mb/s on a 116 ms path, BBR3's start from 12 kB is several round trips (reasoned, not measured). If that is consumer enough, we can measure time to the multiplex's rate after a GOAWAY redial on that path.

(Written by Claude Opus 5.5)

Activity

  1. added
    questBeing tracked/planned in a quest. See `quest/`
    on Nov 12, 2026
  2. kixelated commented on Nov 12, 2026

    @kixelated
    Collaborator

    Thanks, this was a useful consumer and measurement. Planned in #8582:

    • quest/m1/export-order.md: moq-mux's exports all subscribe with the default Order::Descending, so after an outage the relay sends the newest groups first while the one due now expires. By rough arithmetic that alone could explain most of the loss at 11.7 s. The quest measures Ascending against Descending on your outage matrix and on a sustained deficit, then ships the order that loses less video. Descending still has reasons to stay in some cases (recovery to the live edge for ABR players, and not starving short budgets on an aggregated upstream), so it's measured rather than assumed. Your offer to run the matrix against a branch is welcome there.
    • quest/m3/quic-path-return.md: both of your moq-quic findings check out. The fix(quic): send the next probe when a silent peer is heard again #6178 heard-probe is spent during a one-way outage, and the first sample after a two-way outage inflates smoothed RTT. The quest keeps the probe armed, rate-limited, and keeps such a sample out of smoothed RTT. This issue closes with it.
    • quest/m4/quic-deadline.md is restored with your trunks as its consumer. quic-careful-resume stays out until a redial's ramp is measured.

    Anyone who wants to donate tokens can run /quest-start quest/m1/export-order, /quest-start quest/m3/quic-path-return, or /quest-start quest/m4/quic-deadline.

    (Written by Opus 5.5)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    questBeing tracked/planned in a quest. See `quest/`

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions