Repository navigation
quest: the 2026-11-20 relay join regressions are gone on main - #9272
Conversation
Quest-Claim: a9cb4852b362af1c
…out searching beneath it #9202 made Answer::shows build the permit's subtree intersection and rebase the handle's scope for every claim it syncs, though nearly every claim sits at a path both allow. Check that path directly first and keep the search for a claim above a region. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…y-join-regressions # Conflicts: # rs/moq-net/src/lite/publisher/routed.rs # rs/moq-net/src/model/origin.rs
|
No actionable comments were generated in the recent review. WalkthroughThe pull request deletes a Markdown document that described six Relay session benchmark regressions, listed suspected pull requests, and proposed a paired local benchmark run before code changes. Merge Risk:
|
There was a problem hiding this comment.
✅ No new issues found.
Approach: sound — the regressions were already fixed by #9251, so deleting the quest file is the right close-out.
Reviewed changes
- Quest file deleted —
quest/m0/nightly/relay-join-regressions.mdis removed; no code, API, or wire changes.
claude-sonnet-5-5 | 𝕏

Finishes quest/m0/nightly/relay-join-regressions.md by deleting it: the regressions are already gone on
main, so this PR changes no code.Problem
Nightly run 44767045952 flagged six Relay session cells on
moq-lite-07-wip(a759e4cf6againsta4fa80a74). moq-dev/moq#9237 fixed other benches from that window (moq-quic noop waits, stats presence) and did not touch these.Findings
Wall time on this host was too noisy to bisect (load around 22 on 24 cores, ±15% between runs). So the bisect timed each join, and each paced delivery round, in retired user instructions on one pinned core, through a scratch patch to the bench that is not committed. Millions of instructions per iteration:
join_broadcastsrelays=1, 1024join_relaysrelays=8, publishers=16a4fa80a74(base)3b7c9e44995979c7e246059266e(moq-dev/moq#9202)a759e4cf6(the nightly's head)b4496b120(mainwith moq-dev/moq#9251)Answer::shows(rs/moq-net/src/lite/publisher/routed.rs) to search for a covered path both the permit and the handle allow, throughPermit::beneathandorigin::Consumer::reaches_any. Every claim a joining viewer's answer syncs then built a subtree pattern, intersected it with the grant and the limit (each building a pattern index), and rebased the handle's scope. The before/after profile showsmoq_pattern::index::Node::child,Patterns::intersect,Index::contains, and extra malloc/free only after fix(moq-net): lite-07 answers a prefix wider than the grant region by region #9202.Permit::within(...).overlaps). On currentmainevery flagged join cell is back within 2% of base:join_broadcastsrelays=1/2/8 at 64 broadcasts: 8.23, 8.64, 9.48 against base 8.48, 8.88, 9.37.delivery_paced/.../viewers=256is noise. Its nightly rounds were +6.2%, +26.9%, and +36.7%. Its instruction count moved only +1.5% from base toa759e4cf6(viewers=16: +1.6%), and is 29.7 onmainagainst base 29.4. The cell is memory-bound, at IPC 0.35 to 0.45, and its cycles and cache misses varied more between runs of one binary than between base and new.This branch first carried its own fix for #9202's cost: a fast path in
showsthat checks the claim's own path before searching. It recovered the same cells (136.9 and 4.41 above), and the full moq-net, moq-relay, and moq-auth suites passed with an assert that it agreed with the search. #9251 then rewrote the same lines, so the merge ofmaintook #9251's version and dropped the fix as redundant.Impact
Validation
quest checkpasses.Caused by
Answer::showsintersected patterns and rebased the scope for every claim it synced, adding 12 to 20% onrelay_session_join_broadcastsand 14.8% onrelay_session_join_relays(found by check). moq-dev/moq#9251 already removed it.Decisions
relay_session_delivery_paced/.../viewers=256as noise instead of accepting it, since no culprit exists to accept it, as moq-dev/moq#9237 did forrelay_auth_refresh.join_broadcastsinstructions betweena4fa80a74and3b7c9e449. It is under the paired comparison's 10% threshold, and currentmainis within 2% of base anyway.a759e4cf6, is the wall-time check.Follow-ups
None.
Models
Size: M
Implementer: Claude, claude-opus-5-5, medium effort (arm A)
Review
mainonly deletes the quest file. Codex (gpt-6.1-sol, high effort) reviewed the earlier head d256f5d with itsshowsfast path, which the merge ofmaindropped: score 9/10, no findings.(Written by Claude Opus 5.5 with medium)