Repository navigation
Conversation
Frando
force-pushed
the
Frando/send-groups
branch
from
November 16, 2026 11:12
59369d6 to
cf14b6d
Compare
Frando
force-pushed
the
Frando/send-groups-incremental
branch
from
November 16, 2026 11:12
6825b22 to
667b0de
Compare
Frando
changed the base branch from
Frando/send-groups
to
Frando/cleanup-auth
November 16, 2026 13:35
Frando
force-pushed
the
Frando/send-groups-incremental
branch
from
November 16, 2026 13:35
667b0de to
b52f424
Compare
Frando
added a commit
that referenced
this pull request
Nov 16, 2026
…s (merge of #94) Rebuild iroh-live on moq's `main` branch and iroh 1.4, split the code into crates by whether they need iroh, and add an access model so a node decides which peers may connect and what each may read and publish. Every library crate changes its public API, and nothing stays wire-compatible with `main`. The book in `docs/` is rewritten to match. This replaces #87 and contains all of it. ## Why On `main`, iroh-live builds on the crates.io releases of moq, which leaves three problems: - No access control. `Moq::publish` announces every broadcast to every peer that connects, so a call or a room cannot keep its broadcasts to its members. - Two transport adapters. MoQ runs over WebTransport on HTTP/3 on iroh, at an allocation per operation and a copy per write, and a node cannot attach to a MoQ relay. - Duplication. moq `main` now has route selection and failover, rendition selection, A/V sync, capture, codecs and audio output, much of which `moq-media` and `iroh-moq` reimplement. The goal is to use upstream moq for everything it does, keep in iroh-live only what needs iroh, and put the rest in `moq-ext-*` crates with no iroh dependency so each can move upstream on its own. The facade imposes no path layout, endpoint setup or router. ## Crates - `moq-transport-iroh`: a moq-net session over an iroh connection, implementing moq-net's poll transport directly on iroh streams. - `moq-ext-access`: grants, policies, audiences, layouts, and the optional JWT and rcan verifiers. No iroh. - `moq-ext-adaptive`: rendition choice, decoder backoff and handover as pure state machines with no I/O. - `iroh-live-node`: the `Node` with its route table, publications, sessions, relay links, admission, tickets, link stats, `MoqPreset`, `EndpointConfig` and `LiveLayout`. - `iroh-live-rooms`: gossip membership and members-only broadcasts. - `moq-ext-sources`, `moq-ext-publisher`, `moq-ext-player`, `moq-ext-egui`: capture, the encoding `LocalBroadcast`, the decoding `RemoteBroadcast` and `Player`, and egui views. No iroh. - `iroh-live`: the facade, with `Calls`, `SubscriptionExt`, one `Error` and re-exports. - `iroh-live-relay` and `iroh-live-cli` (`irl`): the applications. The media crates know nothing of the network, so one player runs over a direct session, a relay link or another MoQ transport. ## How a broadcast flows A publisher binds an endpoint with `EndpointConfig`, creates a `Node` with a policy, mounts it on its own router, publishes a `LocalBroadcast` at a path, and shares a `BroadcastTicket`. A viewer calls `Node::subscribe_ticket` and plays the result through `SubscriptionExt::remote_broadcast`, which attaches the serving link's measurements so the `Player` adapts with nothing more to wire. `Node::publish` creates the broadcast inside the route table, and sessions read it from there without a copy. Encoders run only while someone reads a rendition. Every session and relay link feeds the viewer's route table, a path resolves to its cheapest route, and moq moves a subscription to another route of the same publisher when its route dies. A node serves only its own publications and never forwards a route it learned from another link. ## Access A policy decides each session: `Policy::decide` returns a `Grant`, two sets of path patterns for subscribe and publish, with an optional expiry, or a rejection. The node decides again on expiry and on `Node::reauthorize`. Policies compose (`Open`, `Closed`, `from_fn`, `filter`, `Queue` for asking the user, `Verify` with `JwtVerifier` or `RcanVerifier`). A layout is the path convention. `LiveLayout` puts broadcasts at `live/<endpoint id>/<name>` and lets each peer publish only under its own id, and `RoomsLayout` does the same under `rooms/<topic>/`. A publication's `Audience` narrows the grant per broadcast: `Granted` reaches every session whose grant covers the path, `Peers(set)` only the named peers. `docs/concepts/access.md` describes the model. ## Calls and rooms A call is a broadcast with `Audience::Peers` over a set that ringing adds to and hanging up removes from. A room is an iroh-gossip topic whose members announce their broadcasts. `Room::state()` returns a watched `RoomState`, and `RoomMember::connection` reports when the last direct session to a member closes. Chat moved out of the library into `irl`. ## Relays The iroh-live relay runs one `moq_relay::Cluster` that browsers reach over WebTransport and iroh nodes over `moq-transport-iroh`. It gets a broadcast by push, when a publisher attaches a relay link, or by pull, when a session names a ticket and the relay dials the publisher. Pulls require a token that grants the ticket's path (or `--open-pulls`), are capped and rate-limited per client, and only work for paths the ticket's publisher owns under a layout. With `--auth-key`, JWTs grant more to iroh clients and browsers alike, and the key reloads on change. A relay link (`Node::attach_relay`) runs moq-net's `Reconnect` with backoff and GOAWAY migration over `https://` or `iroh://`, with tokens that can be replaced on a live link. ## Transport Raw QUIC has no request URI, so the path and token travel in the MoQ SETUP. `export_keying_material` forwards iroh's TLS exporter for moq-lite-07 route key proofs. moq-net opens one send group per subscription, while iroh has one `i32` urgency per stream. `Transport::send_group` returns a `SendGroup` that packs the group's priority and each stream's rank into that urgency, so a group's newest stream goes first and equal-priority groups share bytes round robin. A replacement rendition therefore does not wait behind the old one's backlog, and a switch lands on a link held at its cap. `docs/architecture/transport.md` has the details. After a simultaneous dial, only the oldest session with a peer announces, since moq-lite-06 carries no publisher instance. ## Notable decisions - moq is pinned to `main` at rev `e8075629a` by git, with no `[patch]` block. It includes moq-dev/moq#8561, #8562, #8636, #8639 and #8640. iroh-live cannot publish to crates.io until moq releases from `main`. - Authorization is per session, not per peer, so a decision stays tied to what the peer presented on that session. A dial also asks the policy. - Layouts are conventions, not a fixed scheme: `Node::publish` takes any path, and `BroadcastTicket` holds a full path. - The player samples the link every 200 ms, caps bitrate at the publisher's PROBE delivery estimate, steps down on 10% loss and to the lowest rung on 20%, and holds 500 ms before a downgrade and 4 s before an upgrade. Upgrades decode in parallel and swap once caught up. `MoqPreset` uses BBR3. - With `pipewire`, Linux defaults to the shareable PipeWire camera and the portal screen picker. `irl` enables every backend by default, while `iroh-live` has no default features and a `full` feature. - Config structs are `#[non_exhaustive]` with `Default` and `with_*` builders, so fields can be added without a breaking release. ## Breaking changes Nothing is wire-compatible with `main`; upgrade every node, the relay and the Android app together. - `iroh-moq` becomes `iroh-live-node` and `moq-transport-iroh`, `iroh-rooms` becomes `iroh-live-rooms`, and `moq-media` splits into the `moq-ext-*` crates. - `Live` and `LiveBuilder` are gone in favour of `EndpointConfig`, `Node` and an application-owned router. - Broadcasts move to `live/<endpoint id>/<name>`. `LiveTicket` becomes `BroadcastTicket` (`ilv1:`), room tickets use kind `room2`, and old tickets do not parse. - `VideoTrack`, `AudioTrack` and `MediaTracks` become one `Player`. - `RoomEvent` becomes the watched `RoomState`, `RoomPeer` becomes `RoomMember`, and chat leaves the library. - The relay pulls tickets only for authorized sessions or with `--open-pulls`, and `irl record --latency` is now `--max-age`. docs/migration.md lists every rename and removal with before and after code. ## Testing Beyond what `main` covered, CI now checks the Android JNI crate with cargo-ndk, builds the relay and the facade without media on their own, runs doctests and `cargo doc` with warnings denied, compiles the guide programs in `docs/` as doctests, and checks the book's links. New integration tests cover grants, re-authorization, tokens, relay pull limits and calls, and the patchbay scenario for a switch on a capped link, ignored on `main`, now runs. ## Known limitations - noq 1.4 applies a raised stream priority to queued data only at the stream's next turn. New streams are unaffected, so switches are too. - Sharing between equal-priority groups relies on noq's send fairness, its default. #97 removes that dependency once n0-computer/noq#859 is released. - The relay decides an iroh client once and closes the session on expiry; the client redials with a fresh token. - No transit: a node never forwards another node's broadcast. - Changing a rendition's codec under the same name breaks decoding for viewers. - `rpicam_h264` publishes one hardware-encoded rendition that does not adapt. - The oldest-session rule can go once moq speaks a lite version that carries the publisher instance.
Frando
force-pushed
the
Frando/send-groups-incremental
branch
from
November 16, 2026 15:25
b52f424 to
437eac9
Compare
Patches noq, noq-proto and noq-udp to the head of n0-computer/noq#859, which adds a per-stream incremental flag with RFC 9218 semantics (`SendStream::set_incremental`). The send groups in moq-transport-iroh set it on every stream. The branch is noq 1.5.0, which iroh 1.4's `noq = "1.4"` accepts, and one copy of each crate is in the tree. NOT FOR MERGE until noq#859 lands in a release. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sets RFC 9218's incremental flag on every stream moq-transport-iroh opens or accepts, through `SendStream::set_incremental` from n0-computer/noq#859. The heads of equal-priority send groups then share the bytes round robin even when the connection's transport config turns send fairness off, instead of relying on noq's default. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Frando
force-pushed
the
Frando/send-groups-incremental
branch
from
November 18, 2026 10:20
437eac9 to
eac02eb
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #94. Draft, and not for merge until n0-computer/noq#859 is released.
What it adds
#94 maps moq-net's send groups onto noq stream priorities. The heads of equal-priority groups share one priority, and noq sends streams of one priority round robin, but only while the connection's send fairness is on. That is noq's default, so #94 works on stock noq 1.4. An endpoint whose
QuicTransportConfigsetssend_fairness(false)would get streams of one priority in the order they were written to, and the older of two equal-priority subscriptions would starve the newer one again.This PR sets RFC 9218's per-stream incremental flag (
SendStream::set_incremental(true), added in noq#859) on every streammoq-transport-irohopens or accepts, so the sharing no longer depends on the send fairness setting. The tests for the urgency mapping now also assert that each stream is incremental, and theSendGroupdocs anddocs/architecture/transport.mddescribe the flag.The noq#859 dependency
The first commit patches
noq,noq-protoandnoq-udpto the PR head (danielpetri1/noqfb242c47, version 1.5.0), marked "NOT FOR MERGE: pending n0-computer/noq#859". iroh 1.4'snoq = "1.4"accepts 1.5.0, andcargo tree -i noq-protoshows a single copy. Once noq#859 is released, the patch should be dropped in favor of a version bump.Tests
cargo make check-all(5 check configs, clippy-D warningsin 3 configs, fmt) andcargo fmt --checkpass.cargo nextest run -p moq-transport-iroh --all-features: 21 of 21 pass, includingequal_priority_groups_progress_together.a_switch_lands_while_the_link_stays_capped: 3 of 3 runs pass (6 to 9 s each).scripts/check-links.py: 364 links, none broken.🤖 Generated with Claude Code