moq-net: RAII broadcast announcements via OriginPublish; remove broadcast abort/close by kixelated · Pull Request #1640 · moq-dev/moq

kixelated · 2026-06-06T22:29:20Z

Summary

Makes the origin announcement lifecycle sans-io and reframes what a broadcast is.

Previously publish_broadcast spawned a task per broadcast (web_async::spawn) that awaited broadcast.closed() and then removed it from the origin tree, which required a runtime. Now publishing returns an OriginPublish guard whose Drop does the removal synchronously, so the origin no longer watches closure and no longer spawns. This is the first concrete step toward driving moq-net from poll_xxx/wakers instead of tokio.

Paired conceptual change: a broadcast is just a collection of tracks, so it can no longer be "closed" with a code. BroadcastProducer::abort / BroadcastDynamic::abort are removed. A broadcast ends only when every producer is dropped; pending dynamic track requests are rejected when the last dynamic handler goes away. There is no wire code for broadcast closure, so nothing is lost on the wire. You can now unannounce without closing (drop the guard while the producer keeps serving tracks).

API changes (moq-net, breaking → targeting `dev`)

publish_broadcast returns Result<OriginPublish, Error> (was bool). Hold the guard to stay announced; drop it (or call OriginPublish::unannounce) to remove it.
create_broadcast returns Result<BroadcastPublish, Error>; BroadcastPublish derefs to BroadcastProducer and unannounces on drop.
New Error::Loop (broadcast's hop chain already contains this origin) and reused Error::Unauthorized (path outside allowed prefixes) distinguish the two reject reasons that were previously a bare false/None.
Removed BroadcastProducer::abort and BroadcastDynamic::abort.

The Lock-based origin tree is unchanged — connections stay scoped to their place in the tree so they don't contend on a global lock.

Callers

lite/ietf subscribers: store the guard in the per-broadcast map so Ended / restart / reflected-loop drop it. Side benefit: unannounce-on-Ended is now prompt instead of waiting for the broadcast to fully close.
stats / relay / cli / gst / rtc / boy / examples: hold the guard for the broadcast's lifetime.
FFI edges (libmoq, moq-ffi): keep fire-and-forget by spawning a small boundary watcher that drops the guard when the broadcast closes. Public signatures are unchanged, so the py/swift/kt/go wrappers and doc/lib need no changes.

Cross-package sync

Skipped js/net and doc/concept: this is a Rust-internal lifecycle refactor (RAII guard + removing abort); the wire protocol is unchanged, and the JS @moq/net API has its own idioms. Worth a follow-up if we want JS to mirror the guard/unannounce model.

Test plan

cargo test -p moq-net — 366 pass (origin tests now exercise OriginPublish::drop via a #[cfg(test)] helper that reproduces the old fire-and-forget behavior)
cargo test -p moq-ffi — 18 pass (exercises the relocated announce watcher)
cargo check --all-targets across every changed crate — clean
cargo clippy -p moq-net --all-targets (via nix) — clean
cargo fmt (via nix) applied
full just check — run once any concurrent local moq-cli demo (shared target dir) has exited

🤖 Generated with Claude Code

…st close Make the origin announcement lifecycle sans-io. Previously `publish_broadcast` spawned a task per broadcast that awaited `broadcast.closed()` and then removed it from the tree, which required a runtime. Now publishing returns a guard whose `Drop` does the removal synchronously, so the origin no longer watches closure and no longer spawns. This is paired with a conceptual change: a broadcast is just a collection of tracks, so it can no longer be "closed" with a code. Removed `BroadcastProducer::abort` / `BroadcastDynamic::abort`; a broadcast ends only when every producer is dropped, and pending dynamic track requests are rejected when the last dynamic handler goes away. There is no wire code for broadcast closure, so nothing is lost. API changes (moq-net, breaking): - `publish_broadcast` returns `Result<OriginPublish, Error>` (was `bool`). The guard keeps the broadcast announced; drop it (or call `OriginPublish::unannounce`) to remove it. The announcement is independent of the broadcast's own lifetime, so you can now unannounce without closing the broadcast. - `create_broadcast` returns `Result<BroadcastPublish, Error>`, which derefs to `BroadcastProducer` and unannounces on drop. - New `Error::Loop` (hop chain already contains this origin) and reused `Error::Unauthorized` (path out of scope) distinguish the two reject reasons, instead of a bare `false`. - Removed `BroadcastProducer::abort` and `BroadcastDynamic::abort`. The Lock-based origin tree is unchanged: connections stay scoped to their place in the tree so they don't contend on a global lock. Callers updated to hold the guard for the broadcast's lifetime. The lite/ietf subscribers store it in their per-broadcast maps so Ended/restart/reflected drop it (which also makes unannounce-on-Ended prompt instead of waiting for the broadcast to fully close). The runtime-bound FFI edges (libmoq, moq-ffi) keep fire-and-forget by spawning a small boundary watcher that drops the guard on close, so their public signatures and the language wrappers are unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

kixelated · 2026-06-07T01:12:41Z

+		// Errors only if the origin rejects the publish (loop / out of scope). libmoq origins
+		// are full-scope and broadcasts carry empty hop chains, so this is unreachable; ignore
+		// it to match the prior fire-and-forget contract.
+		if let Ok(publish) = origin.publish_broadcast(path, broadcast.clone()) {


I think we need to store it in the slab? And add an unpublish method.

Also please return an error instead of assuming it will always be impossible.

Done. publish now stores the OriginPublish guard in a new published slab and returns a publish handle; added moq_origin_unpublish (C: moq_origin_unpublish(handle)) which removes the entry and drops the guard to unannounce. It now propagates the origin error via ? instead of swallowing it. Note this drops the old auto-unannounce-on-close behavior: the announcement lives until unpublish, independent of the broadcast closing. (Written by Claude)

kixelated · 2026-06-07T01:14:01Z

-	origin.publish_broadcast(&name, publish.consume());
+	let _publish = origin
+		.publish_broadcast(&name, publish.consume())
+		.expect("origin should allow publishing");


.context instead of .expect for all of these that use anyhow.

Switched all the anyhow call sites (moq-cli client/server, clock/chat/video examples, moq-boy, moq-rtc) to .context("failed to publish broadcast")?. (Written by Claude)

kixelated · 2026-06-07T01:14:23Z

+	// Held for the lifetime of this task; dropping it (on return) unannounces the broadcast.
+	let _publish = origin
+		.publish_broadcast(&settings.broadcast, broadcast_consumer)
+		.map_err(|err| anyhow::anyhow!("failed to publish broadcast {}: {}", settings.broadcast, err))?;


Please just use .context, I don't need to know the name.

Done, now .context("failed to publish broadcast")? with no name. (Written by Claude)

…publish - Remove the loop check (and `Error::Loop`) from `publish_broadcast`. Relays already filter reflections on the announce path, so the origin check was pure redundancy; keep a `debug_assert` as a dev tripwire. `publish_broadcast` now fails only with `Error::Unauthorized` (path out of scope). - libmoq: store the `OriginPublish` guard in a slab and return a publish handle; add `moq_origin_unpublish` to drop it. Propagate the origin error instead of assuming the publish always succeeds. - Use `anyhow::Context` instead of `.expect` / `anyhow!` for the anyhow callers (moq-cli, examples, moq-boy, moq-gst, moq-rtc), dropping the broadcast name from the message. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

kixelated commented Jun 7, 2026

View reviewed changes

kixelated enabled auto-merge (squash) June 7, 2026 01:48

kixelated merged commit 623a36a into dev Jun 7, 2026
2 checks passed

kixelated deleted the claude/interesting-agnesi-609ac5 branch June 7, 2026 01:53

kixelated mentioned this pull request Jun 13, 2026

fix(libmoq): update tests for moq_origin_publish handle return #1710

Merged

4 tasks

kixelated mentioned this pull request Jun 26, 2026

feat(moq-net): add OriginProducer::dynamic + infallible OriginConsumer::request_broadcast #1913

Merged

4 tasks

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

moq-net: RAII broadcast announcements via OriginPublish; remove broadcast abort/close#1640

moq-net: RAII broadcast announcements via OriginPublish; remove broadcast abort/close#1640
kixelated merged 2 commits into
devfrom
claude/interesting-agnesi-609ac5

kixelated commented Jun 6, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

kixelated Jun 7, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

kixelated commented Jun 6, 2026

Summary

API changes (moq-net, breaking → targeting dev)

Callers

Cross-package sync

Test plan

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

kixelated Jun 7, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

API changes (moq-net, breaking → targeting `dev`)