Skip to content

IETF subscribers cannot bound delivery lateness: SUBSCRIBE delivery timeouts are dropped and no max_delay maps from the wire #5881

Closed

Description

@mux-hermoq

An IETF subscriber has no way to bound how late the relay delivers to it. SUBSCRIBE's OBJECT_DELIVERY_TIMEOUT (0x02) and SUBGROUP_DELIVERY_TIMEOUT (0x06) are decoded and dropped (rs/moq-net/src/ietf/subscribe.rs, the _object_delivery_timeout / _subgroup_delivery_timeout bindings, the same on main at 7a13b8f). The IETF serving side then gives every IETF subscriber the largest safe max_delay (serving_subscription in ietf/publisher.rs). doc/concept/standard.md says why: "Delivery timeouts are advisory... Latency enforcement uses presentation time instead of wall-clock delivery deadlines."

That design choice is fine. The problem is that an IETF subscriber has no other way to set its presentation-time budget either. moq-lite subscribers get one through Max Age / max_delay.

Why it matters. Take voice to a listener on a bad link: an agent talking to a person, or a call. The listener wants a gap rather than a growing delay. With no budget, the relay keeps everything queued for the slow subscriber and sends it all, in order, however late.

Measured on moq-relay 0.17.5 (local, default cache settings; the subscriber's link went through moq-shaper --rate 300000 --queue 1s; the publisher sent 1000-byte one-object groups every 20 ms on a track with Timescale 1e6 and a Timestamp on every object):

subscriber's SUBSCRIBE groups delivered lateness p50 / max
no delivery timeout 600 of 600 2.2 s / 11.5 s
OBJECT_ and SUBGROUP_DELIVERY_TIMEOUT = 100 ms 600 of 600 2.3 s / 11.4 s

A 24 s run delivered every group, the last ones about 24 s late. Nothing was reset or skipped, even well past the track's 5 s default max_age. GROUP_ORDER Descending and Ascending gave the same arrival order.

Ask. Give an IETF subscriber a way to set its own max_delay. Mapping SUBGROUP_DELIVERY_TIMEOUT (or OBJECT_DELIVERY_TIMEOUT) on SUBSCRIBE and REQUEST_UPDATE onto the subscription's max_delay would fit the presentation-time model standard.md describes. A group's age would be judged in media time, as for moq-lite, rather than as a wall-clock deadline. The relay would then skip or reset groups that fall further behind the live edge than the subscriber can use.

Until then, the client works around it (muxinc/relay-sdks): when audio keeps arriving too late to play, it cancels the subscription and resubscribes at the live edge, which throws the relay's backlog away.

Related: #712 and #856 (closed delivery-timeout PRs). For FETCH, main already keeps object timestamps (#4822, #5494); 0.17.5 sends FETCH objects with no properties.

Activity

  1. added
    questBeing tracked/planned in a quest. See `quest/`
    on Oct 19, 2026
  2. kixelated commented on Oct 19, 2026

    @kixelated
    Collaborator

    Decision: a new parameter rather than the IETF delivery timeouts. OBJECT_DELIVERY_TIMEOUT resets the newest group too (a late joiner's catalog would be cut), so neither timeout matches moq-lite's max age. The new SUBSCRIBE/REQUEST_UPDATE parameter lives in draft-lcurley-moq-timestamp, is negotiated with its own Setup Option, and maps onto max_delay: only groups superseded by a newer one expire, judged on media time. The IETF timeouts stay dropped, and standard.md will say so.

    Tracked by quest/m1/ietf-max-age.md. Your client would send the new parameter.

    Planned in #5969. Anyone who wants to donate tokens can run /quest-start quest/m1/ietf-max-age to pick it up.

    (Written by Claude 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