Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions doc/concept/standard.md
Original file line number Diff line number Diff line change
Expand Up @@ -28,13 +28,13 @@ subset drops. What a peer actually observes against this implementation:
- **Pull.** Subscribers ask. Single-track `PUBLISH` offers are declined, unless they answer our `SUBSCRIBE_TRACKS` (below); announce a namespace and serve the resulting subscriptions. Announcements go out unsolicited. A client asks only for the prefixes its readers watch, and a server or relay link for every prefix it may discover. We never hold two overlapping requests at a peer that did not declare solicit: a broader one replaces the narrower ones once they are closed. The solicit `SETUP` option makes us wait to be asked. Without it, a peer also gets each match as a `NAMESPACE` on its `SUBSCRIBE_NAMESPACE` stream, so it hears the namespace twice. We send `PUBLISH` only to answer a `SUBSCRIBE_TRACKS` (below).
- **`SUBSCRIBE_TRACKS` lists, it never pushes.** It lists every track of every broadcast under its prefix, later ones included. With `FORWARD=0`, each track gets a paused `PUBLISH` carrying its Track Properties, and a `PUBLISH_DONE` once the track is no longer listed: `TRACK_ENDED` if it finished, even just before its broadcast ended, and an error otherwise, since the track may go on elsewhere. A `REQUEST_UPDATE` with `FORWARD=1` resumes one, served from the cache like a `SUBSCRIBE` with the update's filter and answered with `LARGEST_OBJECT`; `FORWARD=0` pauses it again. Each paused `PUBLISH` opens only as the peer's stream credit admits it, so a peer that grants none holds none. A listing that omits `FORWARD` or sets it to 1 asks for every track's objects: on draft 22 each track gets `PUBLISH_SKIPPED`, and the subscriber `SUBSCRIBE`s to what it wants; draft 18 has no such skip, so the request is refused with `NOT_SUPPORTED`. A prefix overlapping another listing gets `PREFIX_OVERLAP`, a `TRACK_PROPERTY_FILTER` gets `NOT_SUPPORTED`, and a `REQUEST_UPDATE` on the listing itself ends it. On draft 18, a `SUBSCRIBE` to a track we `PUBLISH`ed gets `DUPLICATE_SUBSCRIPTION`. Cancelling the listing stops new `PUBLISH`es, while those sent live on. A broadcast is listed only while our publish grant covers it, from scratch once it does. Rust sends it upstream for a listing (below); JavaScript does not.
- **A Rust TRACKS listing sends `SUBSCRIBE_TRACKS` upstream, only on demand.** Listing a broadcast from a moq-transport peer sends `SUBSCRIBE_TRACKS` with `FORWARD=0` while someone lists it, never on connect, announce, or subscribe, so a peer that refuses, resets, or ignores it affects only that listing. Each listed track's paused `PUBLISH` is its upstream copy, which a reader resumes rather than subscribing again. Only a `PUBLISH_DONE` with `TRACK_ENDED` removes a listed track, and it lists as finished; any other end, including the error our own answerer sends for a track that left unfinished, fails the listing, and readers list again.
- **History is fetched in ranges.** In Rust, Object ID gaps must be marked by Prior Object ID Gap properties for a group to be fetchable. A `FETCH` returns a range of groups from the cache, which a relay fills upstream with one `FETCH` per run of missing groups, or on draft 18 the saved prefix of the group a new subscription just joined. The range ends at Largest Object's group and reads it whole, so a `FETCH` of a group still being written waits for it to finish. That departs from draft 22's cap at Largest Object: a cache cannot know an open group's real Largest Object, and a cut group would be cached downstream as complete. On draft 18, `FETCH_OK` End Object 0 covers the entire End Group, matching the request; Rust also accepts the legacy answer naming the next group with Object 0. On draft 22, `FETCH_OK`'s End Location is the requested end, capped at Largest Object's group: a range holding no objects gets an empty fetch stream, and only a start past Largest Object is refused, with `INVALID_RANGE`. A Rust relay retains the upstream Largest Object from `TRACK_STATUS`, so even a fetch-only copy can resolve a draft 22 `LOCATION_FILTER` relative to it; without a publisher report or live feed, it refuses that filter as `NOT_SUPPORTED`. JavaScript publishing answers a `FETCH` from its cache alone, and a joining `FETCH` only for the head of its subscription's own group. It honors `FILL_TIMEOUT` on a `FETCH` or a fill as a total budget for waiting on objects that have not arrived, zero meaning cache only; once it runs out, the rest of each group still arriving is marked an Unknown gap on draft 18 and a Timed-Out gap on draft 22, never left implying the objects do not exist. It bounds the wait, not how old a cached object is. Rust still refuses `FILL_TIMEOUT` with `NOT_SUPPORTED`. A JavaScript subscription joins at the current group: on draft 18 it fetches the group's head with a Relative Joining `FETCH`, and on draft 22 asks for it with a fill. Datagrams are never fetchable. A Rust reader that only fetches learns the track from `TRACK_STATUS` instead of subscribing, so a finished track stays fetchable; a reader whose publisher refuses `TRACK_STATUS` as `NOT_SUPPORTED` still subscribes.
- **History is served one group at a time.** Our moq-transport publishers answer a `FETCH` only within one group, a draft-18 joining `FETCH` for the subscription's current group head, and a draft-22 fill for the largest group's head. Whole-group fetches wait for an open group to finish. Both implementations accept and ignore `FILL_TIMEOUT`, including zero. A JavaScript subscription joins at the current group: on draft 18 it fetches the group's head with a Relative Joining `FETCH`, and on draft 22 asks for it with a fill. Datagrams are never fetchable. Subscribers still request ranges upstream, one per run of missing groups; a third-party publisher can answer those ranges and mark their gaps. A Rust reader that only fetches learns the track from `TRACK_STATUS`, so a finished track stays fetchable; a reader whose publisher refuses `TRACK_STATUS` as `NOT_SUPPORTED` still subscribes.
- **Compression.** Rust peers exchange compressed tracks when both offer the [Flate extension](/draft/moq-flate). Other peers receive one shared inflated copy. A publisher that seals the compression declaration makes that fallback unavailable: an incompatible peer is refused, preserving the sealed metadata. JavaScript does not offer the IETF extension.
- **Timing.** A track whose `SUBSCRIBE_OK` declares no `TIMESCALE` is untimed. A Rust reader that never subscribed takes the units from `TRACK_STATUS`.
- **Strict SETUP and parameters.** Repeated unknown `SETUP` options, GREASE included, are accepted if well-formed; a repeated known option is rejected. A `GROUP_ORDER` other than Ascending (1) or Descending (2) closes the session. An unknown message parameter, or one defined only for another message, closes the session. Other malformed control input closes the session with the draft's code too, such as a server `SETUP` carrying `PATH` or `AUTHORITY`, a namespace field ending in `/` (it would map to a path with an empty segment), a request cut short by FIN, or a second `GOAWAY` on one stream; a first `GOAWAY` on a request stream is legal.
- **One `SETUP` credential**, forwarded to the [auth server](/bin/relay/auth#the-contract) unverified. A second `SETUP` token closes the session. We advertise no token cache, so an alias reference in `SETUP` closes it with `PROTOCOL_VIOLATION`, one on a request with `UNKNOWN_AUTH_TOKEN_ALIAS`, and a registration on a request with `AUTH_TOKEN_CACHE_OVERFLOW`. Where the Rust side authorizes the peer or verifies tokens, a token on a `SUBSCRIBE` or `PUBLISH_NAMESPACE` is verified like an in-band session token and authorizes that one request alone, even where that authorization already covers it; it never widens the session. It is refused with `NOT_SUPPORTED` when nothing verifies it, when it is empty, or when its Token Type is anything but 0 (out of band). A side that does neither, JavaScript included, ignores the token. Rust refuses a request carrying more than one distinct token with `NOT_SUPPORTED`, since it verifies one; JavaScript, which verifies none, ignores them all. A token-authorized `SUBSCRIBE` still stays within the grant the peer gave us. A token-authorized `PUBLISH_NAMESPACE` needs a grant covering everything beneath the namespace, and the tracks we subscribe under it rely on that announce instead of our limit, ending with it. An update renews the token, and a refused renewal ends only that request. Rust and JavaScript send the token a route names (`Route::auth`) on its `PUBLISH_NAMESPACE`, and swapping it renews the request with `REQUEST_UPDATE`; the inline `NAMESPACE` that answers a solicited `SUBSCRIBE_NAMESPACE` carries no token. A token added to a reconnecting connection's auth handle rides every session it makes, so the announcement keeps it across a reconnect; one added to a single session's handle ends with that session. A token on any other request is decoded and grants nothing, and a request repeating one is refused with `MALFORMED_AUTH_TOKEN`. The negotiated [MoQ Auth extension](/concept/moq-lite#authorization) accepts additional session tokens in band.
- **Pause delivery.** `FORWARD=0` on a `SUBSCRIBE` or its update drops the subscription, closing its data streams and holding nothing upstream, while the request and its slot stay. `FORWARD=1` subscribes again as a new `SUBSCRIBE` with the same filter would, so nothing carries across the pause: its `REQUEST_OK` names the live edge it joined at. A paused update may change only priority and max age.
- **Refused, not fatal.** A legal request this stack does not serve is rejected on its own and the session stays up: the fetch forms above get `NOT_SUPPORTED`. Range Filters get `INVALID_FILTER` on draft 22 because the advertised filter limit is zero. On draft 22, namespaces whose first tuple field is exactly `.` or `.session` get `DOES_NOT_EXIST` before application lookup; other dot prefixes still reach the application. Subscription updates accept priority, negotiated subscriber max age, and delivery timeouts. Rust also carries range changes, keeps bounded subscriptions updatable, renews request tokens, and judges `NEW_GROUP_REQUEST` as a join (see dynamic groups below). On draft 22, widening fills cached history on a separate fetch stream, completing earlier groups and handing the newest head to its live subgroup tail; fills follow the requested group order. Other supported-but-unimplemented preferences end that subscription; a `TRACK_PROPERTY_FILTER` there closes the session with `PROTOCOL_VIOLATION`, since it belongs only to `SUBSCRIBE_TRACKS`. A namespace update may change its prefix; suffixes after the acknowledgement are relative to that new prefix.
- **Refused, not fatal.** A legal request this stack does not serve is rejected on its own and the session stays up: multi-group `FETCH` and joining requests reaching before the joined group get `NOT_SUPPORTED`. Range Filters get `INVALID_FILTER` on draft 22 because the advertised filter limit is zero. On draft 22, namespaces whose first tuple field is exactly `.` or `.session` get `DOES_NOT_EXIST` before application lookup; other dot prefixes still reach the application. Subscription updates accept priority, negotiated subscriber max age, and delivery timeouts. Rust also carries range changes, keeps bounded subscriptions updatable, renews request tokens, and judges `NEW_GROUP_REQUEST` as a join (see dynamic groups below). On draft 22, a wider or historical fill fails on its separate fetch stream while its subscription stays open; a largest-group fill hands its head to the live subgroup tail. Other supported-but-unimplemented preferences end that subscription; a `TRACK_PROPERTY_FILTER` there closes the session with `PROTOCOL_VIOLATION`, since it belongs only to `SUBSCRIBE_TRACKS`. A namespace update may change its prefix; suffixes after the acknowledgement are relative to that new prefix.
- **Objects and datagrams.** Only subgroup 0 is accepted: a stream for any other subgroup, whether its header gives the ID or takes it from the first Object ID, is stopped and the session stays up. Rust relays preserve Object IDs, gaps, object properties, FIRST\_OBJECT, and END\_OF\_GROUP status objects between moq-transport peers for subgroup 0, including across a moq-lite-07 hop between two relays that negotiate the [ietf](/draft/moq-ietf) extension. Datagrams retain their IDs, properties, priority, and status. Plain Rust readers and moq-lite peers without the extension see consecutive frames and only normal datagram payloads at Object ID 0; other datagram objects and statuses are dropped on those surfaces. JavaScript supports only normal datagrams at Object ID 0. A malformed datagram closes the session; a relayed one the downstream draft cannot carry is dropped. A relayed Timestamp leaves in the units the downstream response declares, and is dropped where it declares none, as for a plain frame; an immutable one that would change or vanish is refused. On a timed track the frame holds each object's time, so a mutable Timestamp that only repeats the immutable one is not restated; a status object has no frame, so it keeps its Timestamp among its properties. Only property 0x10 is read as a Timestamp. Property 0x06 stays opaque and keeps its value regardless of timestamp units; it is dropped between drafts 18 and 22, or refused when immutable, because draft 22 assigns it to SUBGROUP\_DELIVERY\_TIMEOUT. A status object carrying properties is refused. A `FETCH` has no status objects, so its response ends at the last frame. A group cannot resume across route changes by frame ordinal once Object IDs skip ahead of the ordinals on either route; recovery refuses it instead of splicing different objects.
- **Dynamic groups.** A track advertises `DYNAMIC_GROUPS` when its publisher sets a refresh ratio, and a track received with `DYNAMIC_GROUPS` gets a ratio of 0. A `NEW_GROUP_REQUEST` counts as one more subscriber joining, which a `SUBSCRIBE` already is: it asks for a new group only when the newest group's bytes after its first frame exceed the ratio times the first frame. Joins coalesce until the asked-for group arrives. A Rust relay sends `NEW_GROUP_REQUEST` upstream, on the `SUBSCRIBE` or in a `REQUEST_UPDATE`, once a join wants a group past the newest it holds; a relay holding nothing yet asks for none. On a track without a ratio, an update carrying one ends the subscription with `UPDATE_FAILED`. A moq-lite-07 hop (`moq-lite-07-wip`, opt-in) carries the same demand natively: the ratio in `TRACK_INFO` and the asked-for group in `SUBSCRIBE_UPDATE`. Earlier moq-lite versions carry neither, so a track received over one never refreshes.
- **Priority.** Higher is served first. This stack's priorities run from -128 to 127 with 0 unset, and go out on the IETF wire, which ranks lower first, as `127 - priority`: an unset priority is IETF 127. An omitted subscriber or publisher priority reads as the drafts' default of IETF 128, one step below unset, so a relay forwards it unchanged. A relay's upstream subscription carries the highest priority its subscribers ask for and updates it as they come and go; its group order stays as the `SUBSCRIBE` asked, since no draft lets an update change it.
Expand Down
4 changes: 2 additions & 2 deletions js/net/bench/fills.ts
Original file line number Diff line number Diff line change
Expand Up @@ -64,7 +64,7 @@ for (const groups of [2, 16, 128]) {
const id = BigInt((round + 1) * 2);
const parameters = new Parameters();
parameters.fillParameters = Filter.encodeFill(
{ filter: { kind: "unfiltered" }, groupOrder: "ascending", rangeFilters: false },
{ filter: { kind: "relative", groups: 1n }, groupOrder: "ascending", rangeFilters: false },
version,
);
await client.writer.u53(SubscribeUpdate.id);
Expand All @@ -77,7 +77,7 @@ for (const groups of [2, 16, 128]) {
if ((await reader.u53()) !== FetchHeader.type) throw new Error("wrong fill stream");
if ((await FetchHeader.decode(reader, version)).requestId !== id)
throw new Error("wrong fill ID");
if ((await reader.readAll()).byteLength < groups) throw new Error("short fill");
if ((await reader.readAll()).byteLength === 0) throw new Error("short fill");
}),
);
if (round > 1) samples.push((performance.now() - start) * 1000);
Expand Down
Loading
Loading