Repository navigation
feat(moq-net,net)!: lite-07 FETCH drops Start Byte - #8837
Conversation
Quest-Claim: 92cdb4d94a202162
Backs out #8699. A failover mid-frame refetches the frame whole and skips the bytes already read, as before. Frames are capped at 32 MiB, so the byte offset saved at most one frame's prefix per failover, on one hop, and only for chunk readers. The draft, Rust, and JS drop the field together; lite-07 is unpublished. The 5 MiB failover integration test stays as half_read_frame.rs and now asserts the new route resends the whole frame. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
No actionable comments were generated in the recent review. WalkthroughFETCH no longer includes a byte offset within its starting frame. Frame and group storage, request routing, and recovery now use frame positions. Publishers serve complete frames, and consumers skip bytes already read after recovery. The changes also update protocol drafts, quest documents, and tests. Merge Risk:
|
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
This review covers the full backout of #8699's FETCH Start Byte across the drafts, moq-net, @moq/net, and the quest files.
- Wire and drafts: lite-07-wip FETCH drops
Start Bytein both the encoder and decoder, in Rust and JS. The lite draft loses the field, its Positions sentence, and its changelog line. The ietf draft loses its non-zero Start Byte ban. - Group model:
group::Bounds::byteand the wholeLeadpath are gone (create_lead,lead_at,Consumer::at,poll_next_frame_at, the lead's billing and shedding intail/trim).boundgoes back to a plainstart_at. - Failover read:
frame::Consumer'sRecoverasks for the frame by index alone, reads the replacement withpoll_next_frame, and keeps its ownread_idxagainst the new whole-framesource. Withoutbase, the offset logic inpoll_read_chunk/poll_read_allgoes back to plainread_idx. - Fill and fetch plumbing: the byte is removed from
Ask,Pending,Inner,GroupFill,Fetching, andgroup::Fetch. Fill merging compares frames only. - Publisher and subscriber serving: the Rust fetch serve goes back to
with_range(start_frame..)with theindex() != start_frame→Laggedguard, andbuffer_frame_infoalways announcessize.FrameIngest::dispatchno longer routes to a lead. JS#runFetchGroupwrites whole frames throughsendPayload. - Regression test:
half_read_frame.rs(formerlybyte_holes.rs) now checks that the reader gets the 5 MiB frame byte-for-byte and thatBresends the whole frame. - Quests:
fetch-start-byte.mdis deleted,byte-holes.mdrecords the backout and a maintainerWaitingitem, and the links inlite07-finalize.mdandyield-relay-links.mdare updated.
A repo-wide search found no leftover start-byte or Lead references in the code, drafts, doc/, or the TLA+ models.
claude-opus-5-5 | 𝕏
|
Maintainer decision (2026-11-15, /quest-complete):
(Written by Opus 5.5) |
Maintainer decision on #8837: no consumer sends frames over 32 MiB, so mid-frame resume is out until one does. JS byte holes only mirrored it. The unbounded-groups epic keeps the Rust frame cap, and compressed-stream no longer waits on lifting it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
This review covers the delta since the prior Pullfrog review (853a4a1 → abe4fb7), which is quest documentation only.
- Deleted byte-hole quests:
byte-holes.mdandjs-byte-holes.mdare removed, and no links to them remain. - Epic rescoped:
quest/m3/unbounded-groups/README.mdbecomes "Unbounded groups", keeps Rust frames capped at 32 MiB, and records the maintainer decision. - Compressed stream:
compressed-stream.mddrops itsRequiredlink to Byte holes and tests at the plain cap in Rust. - Wording:
time-addressing/model.mdno longer names byte-holes as the reason for#[non_exhaustive]onPosition.
claude-sonnet-5-5 | 𝕏
# Conflicts: # quest/m2/time-addressing/model.md # quest/m2/yield-relay-links.md # rs/moq-net/src/lite/subscriber.rs
|
Applies the maintainer decision as of e5f4e26: (Written by Claude Opus 5.5 with medium effort) |
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
This review covers the delta since the prior Pullfrog review (abe4fb7 → e5f4e26): a merge of origin/main (f3f4cc5) and how it was resolved.
- Merge resolution in
FrameIngest::header: kept main's continuation clause(!records.relayed && id == self.index)and dropped onlygroup.lead_at(id). A continuation that re-sends plain frames its group already holds is still accepted. FrameIngest::dispatch: dropped main'slead_at/create_leadearly return, so a frame below the held count goes straight to theResume/Skippath. This matches the refetch-and-skip model.- Quest files: dropped this PR's edits to
time-addressing/model.mdandyield-relay-links.md, since main finished and deleted both. After the merge, no links tofetch-start-byte,byte-holes, oryield-relay-linksremain.
A repo-wide search after the merge found no leftover lead_at, create_lead, start_byte, startByte, hasByteStart, sendPayloadFrom, or poll_next_frame_at references.
claude-opus-5-5 | 𝕏
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
This review covers the delta since the prior Pullfrog review (e5f4e26 → b61d02d): one merge of origin/main with no conflicts.
- Merged
origin/main: the merge changes only nearby context lines inquest/m2/lite07-finalize.md, where main added a lite-07 time-addressing link and removed the ROUTE's RTT line. This PR still only removes thefetch-start-byteline there.
After the merge, a search of the whole repo found no lead_at, create_lead, start_byte, startByte, hasByteStart, has_byte_start, sendPayloadFrom, or poll_next_frame_at references. It also found no links to fetch-start-byte or byte-holes.
claude-opus-5-5 | 𝕏

Finishes quest/m2/fetch-start-byte.md and deletes it.
Summary
Backs out moq-dev/moq#8699, as the maintainer picked on moq-dev/moq#8764. lite-07-wip FETCH no longer carries
Start Byte. A frame read partway when its route fails is refetched whole from the next route, and the reader skips the bytes it already read (frame::Consumer'sRecover), as before #8699.Frames are capped at 32 MiB (
MAX_CACHE_BYTES), whole-frame reads asked byte 0 anyway, and only chunk readers saved at most one frame's prefix on one hop per failover. That did not pay for aLeadin the group model, the start-byte plumbing through the fill, subscriber, and publisher in Rust and JS, and a field frozen in lite-07.Removed:
Start Bytein the lite draft's FETCH, its Positions sentence, and its changelog line; the ietf draft's ban on a non-zero Start Byte for a track with a Version, and its changelog line.moq-net):lite::Fetch::start_byte,Version::has_byte_start,group::Bounds::byte, theLead(create_lead,lead_at,group::Consumer::at,poll_next_frame_at), the fill's byte (Ask::byte,Pending::byte,GroupFill::byte_start),frame::Consumer::base, and the publisher's serving from a byte.@moq/net):Fetch.startByte,hasByteStart,sendPayloadFrom, and the publisher's serving from a byte.The revert of the squashed #8699 commit conflicted only where later PRs renamed
sequencetokeyor touched the same tests; those were resolved to the current names.Merging main (2026-11-15) conflicted in
rs/moq-net/src/lite/subscriber.rs'sFrameIngest::header, where main let a continuation re-send frames its group holds (!records.relayed && id == self.index); the resolution keeps that clause and drops only the removedlead_atcheck.quest/m2/time-addressing/model.mdandquest/m2/yield-relay-links.mdfinished on main, so their edits here are gone. After the merge:cargo test -p moq-net --all-features(3388 unit, 606 integration) andjust checkpass. A second merge of main (conflict-free) was followed bycargo test -p moq-net --all-features(3373 unit, 606 integration),bun test src/liteinjs/net(899 pass),just check, andquest check, all passing.Quests: Byte holes (
quest/m3/unbounded-groups/byte-holes.md) and JS byte holes (quest/m3/unbounded-groups/js-byte-holes.md) are deleted (see Decisions). The unbounded groups epic keeps Rust frames capped at 32 MiB and records the decision; Compressed stream no longer waits on lifting the plain cap. Lite-07 finalize drops its link, and Yield on relay links no longer points at Start Byte as its example.Decisions
quest/m3/unbounded-groups/byte-holes.md(rejected: keep it parked). No consumer sends frames over 32 MiB, so mid-frame resume is out until one does.quest/m3/unbounded-groups/js-byte-holes.mdtoo (rejected: keep it). It only mirrored Byte holes for@moq/netand required it, so the same reason applies.Requiredlink to Byte holes: in Rust it removes only the compressed-slice guard, keeps the inflated size checked against the plain cap, and tests at that cap; JS still goes past 32 MiB, since JS plain frames are already uncapped.Choices made without asking
rs/moq-net/tests/integration/half_read_frame.rs(it wasbyte_holes.rs). It now asserts the reader gets the whole frame correctly and the new route resends the whole frame, so refetch-and-skip keeps an over-the-wire regression.resumehalf-read tests. The existingresumetest where the replacement holds the frame whole and the reader continues withcdcovers refetch-and-skip at the model level.Maintainer:Waitingitem (name a consumer of frames over 32 MiB) rather than prose, so it surfaces inquest waiting.Public API and wire impact
@moq/net'sFetchlosesstartByte(a lite message class, not used outside the package). Rust's surface was crate-private.Start Byte, breaking for lite-07-wip peers built since feat(moq-net,net)!: lite-07 FETCH Start Byte continues a half-read frame by byte #8699; Rust and JS change together. Published versions are unchanged. No migration: lite-07 is unpublished.Validation
half_read_frame::a_half_read_frame_refetches_whole_and_skips_what_was_read: relaysP -> A, B -> Eover lite-07-wip.Ereads 2 MiB of a 5 MiB frame throughA,A's session dies, andEreads the rest throughBbyte-for-byte.B's session sent between the frame's full size and 1 KiB more, so the frame was refetched whole and the read part skipped.cargo test -p moq-net --all-features: all pass (3375 unit, 602 integration).cargo clippy -p moq-net --all-features --all-targets -D warningsand the--no-default-features --libbuild: clean.just check: passes (Rust lint across the workspace, JS, drafts, markdown, quest).just test interop --all: 211 passed, 0 failed.Caused by
Start Byteand aLeadfor a saving the 32 MiB frame cap makes marginal (found by reviewer: the feature review)Follow-ups
None. Byte holes and JS byte holes are deleted rather than left parked.
Models
Size: S
Implementer: Claude Code, claude-opus-5-5, unknown effort (arm A)
Implementer: Claude Code, claude-opus-5-5, medium effort (iterate, S, arm A)
(Written by Claude Opus 5.5 with medium effort)
🤖 Generated with Claude Code
Review
Rewritten by
quest reviewafter each round.Head: e5f4e26
Base: origin/main at f3f4cc5
Risk: high,
The overall branch removes wire-protocol fields and public API, potentially breaking dependent implementations whose rollback would require coordination.Classified: medium by default
Reviewer: Codex, gpt-6.1-sol, medium effort
Session:
01a1df97-1c5b-7573-8955-40c36b2ca597, round 3Score: 9/10,
The merge preserves upstream continuation handling while removing byte-offset support consistently, with no actionable defects found by source inspection; builds and tests were not run as instructed.Raised the risk to high:
The overall branch removes wire-protocol fields and public API, potentially breaking dependent implementations whose rollback would require coordination.Rounds: 3; first round 0; findings 0: fixed 0, refuted 0, open 0
Findings: none