Repository navigation
fix(congestion): preserve low-rate pacing debt and retune BBR3 - #7606
Merged
Merged
Conversation
Quest-Claim: 403dc0057011512e
kixelated
changed the title
Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
Co-Authored-By: GPT-6.1 Sol <noreply@openai.com>
The full local gate now passes, so the pacer correction and BBR3 retune need no further work under quest/m1/pacer-low-rate.md. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Contributor
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 9 minutes. View limit details
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. |
kixelated
commented
Nov 2, 2026
kixelated
left a comment
kixelated
left a comment
Collaborator
Author
There was a problem hiding this comment.
Automated review by review (OpenAI)
Reviewed commit: dc82437
No actionable bugs found in the changed areas. Carrying overshoot into the shared pacer's next refill addresses the lost-credit root cause in both pacing paths; the capacity-change clamp preserves bounded repayment for unpaced probes. The existing BBR3 quantum mechanism is a proportionate place to reduce low-rate aggregation, without adding API/wire surface or compensating with higher gains. The deterministic one-/two-datagram sweep and first-RTT coverage target the relevant boundaries.
Verification: static review of the diff, pacer accounting/tests, BBR3 quantum/offload integration, and moq-quic's pacing/batch caller. I did not run tests or reproduce the PR's reported benchmark/CI results; the competition measurements remain author-reported evidence.
(Written by OpenAI)
![pullfrog[bot]](https://avatars.githubusercontent.com/in/1768019?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
✅ No new issues found.
Reviewed changes
- Pacer debt accounting (
rs/moq-congestion/src/pacer.rs) —on_transmitnow records the bytes sent past available credit in adebtfield, bounded to one burst, andrefillrepays that debt before adding earned credit. Bothdelayanddelay_at_raterequiredebt == 0and folddebtinto the wait. This charges the second datagram the transport releases on one packet of credit at low rates, instead of discarding its cost through saturating subtraction. - BBR3 quantum retune (
rs/moq-congestion/src/bbr3/mod.rs) — the initialsend_quantumand theset_send_quantumfloor drop from2 * smsstosmss, soburst()yields one-datagram aggregates below roughly 19 Mbit/s. Draft §5.6.3's default floor ismax(..., 2 * C.SMSS)but the section explicitly permits alternatives and directs implementations to use the smallest feasible quanta. One coupled effect the description doesn't mention:offload_budget = send_quantum + delayed_ack_termalso drops bysmss, which only binds when a connection negotiatesack_eliciting_threshold >= 4(inert under the default threshold). - Tests — new
repays_a_packet_sent_past_the_burst_budgetandcapacity_changes_bound_unpaced_debt; the throughput sweep now spans 1_250..12_500_000 B/s over 10 s with one- and two-datagram batches against the initial-credit-plus-final-batch bound; first-RTT calibration extends through long-RTT low-rate cases. - Quests — deletes the finished
quest/m1/pacer-low-rate.mdand updates them1andquic-realtimeREADMEs.
Verified the debt invariants (tokens <= capacity; debt > 0 ⇒ tokens == 0) across both pacing paths and their capacity-change clamps: no stall, no underflow/overflow, and the paced path is bounded to one datagram of overshoot. Unpaced loss probes can leave more than one burst unrepaid, but that is the documented hostile-peer bound, strictly more conservative than the pre-fix code, and externally bounded by the window and send-rate cap. Two independent specialist passes on the pacer accounting and the BBR3 quantum integration found no violation.
deepseek-v4.1-flash (free via Pullfrog for OSS) ｜ 𝕏
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
github-merge-queue
Bot
removed this pull request from the merge queue due to failed status checks
Nov 2, 2026
Contributor
|
quest land watch: the merge queue removed this pull request at 4d1974a because its The land fixer works on it in the background, and the watcher re-enqueues it at the back of the queue once it is fixed or judged flaky, or hands it back here with what it needs.
…0.032s] (1774/6817) moq-gst source::imp::session_tests::malformed_input_is_not_a_lost_source
PASS [ 0.289s] (1775/6817) moq-gst source::imp::session_tests::a_relisted_rendition_keeps_its_pad
PASS [ 0.067s] (1776/6817) moq-gst source::imp::session_tests::keylog_reaches_the_client
PASS [ 0.052s] (1777/6817) moq-gst source::imp::session_tests::rendition_subscribes_on_its_referenced_broadcast
PASS [ 0.200s] (1778/6817) moq-gst source::imp::session_tests::a_sibling_rendition_relisted_while_draining_does_not_replay_its_finish
PASS [ 0.177s] (1779/6817) moq-gst source::imp::session_tests::an_unserved_rendition_does_not_claim_the_first_pad_name
PASS [ 0.198s] (1780/6817) moq-gst source::imp::session_tests::an_ended_broadcast_resumes_on_the_same_pad
PASS [ 0.123s] (1781/6817) moq-gst source::imp::session_tests::paused_to_ready_returns_after_the_connection_closes
FAIL [ 0.108s] (1782/6817) moq-gst source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad
stdout ───
running 1 test
test source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad ... FAILED
failures:
failures:
source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad
test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 170 filtered out; finished in 0.07s
stderr ───
thread 'source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad' (1245554) panicked at rs/moq-gst/src/source/imp.rs:3788:13:
assertion `left == right` failed: [Pad("video_0"), Segment("video_0", 79000), Buffer("video_0", 1), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 94000), Buffer("video_0", 2), Buffer("video_0", 3), Buffer("video_0", 4), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 94000), Buffer("video_0", 2), Buffer("video_0", 3), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 62000), Buffer("video_0", 5), Buffer("video_0", 6), Buffer("video_0", 7), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 62000), Buffer("video_0", 5), Buffer("video_0", 6), Buffer("video_0", 7), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 73000), Buffer("video_0", 8), Buffer("video_0", 9), Buffer("video_0", 10), Start("video_0"), Caps("video_0", "video/x-h264"), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 2809000), Buffer("video_0", 11), Buffer("video_0", 12), Buffer("video_0", 13), Eos("video_0")]
left: 1
right: 0
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
Cancelling due to test failure: 23 tests still running
PASS [ 0.051s] (1783/6817) moq-gst source::imp::session_tests::the_source_retries_without_a_budget
PASS [ 0.049s] (1784/6817) moq-gst source::imp::tests::aac_caps_follow_the_description_framing
PASS [ 0.136s] (1785/6817) moq-gst source::imp::session_tests::stop_from_a_runtime_worker_does_not_deadlock
PASS [ 0.137s] (1786/6817) moq-gst source::imp::session_tests::stop_from_its_own_reader_does_not_deadlock
PASS [ 0.042s] (1787/6817) moq-gst source::imp::tests::an_unrepresentable_timestamp_is_refused
PASS [ 0.202s] (1788/6817) moq-gst source::imp::session_tests::audio_and_video_share_a_run_timeline_across_scales_and_restarts
PASS [ 0.135s] (1789/6817) moq-gst source::imp::session_tests::stop_inside_another_executor
PASS [ 0.121s] (1790/6817) moq-gst source::imp::session_tests::stop_returns_after_the_connection_closes
PASS [ 0.125s] (1791/6817) moq-gst source::imp::session_tests::stop_releases_a_blocked_push
PASS [ 0.643s] (1792/6817) moq-gst source::imp::session_tests::a_delisted_sibling_rendition_follows_no_later_instance
PASS [ 5.248s] (1793/6817) moq-cli::integration import::a_short_import_drains_its_media_to_a_waiting_subscriber
PASS [ 0.379s] (1794/6817) moq-gst source::imp::session_tests::an_update_keeps_the_run
PASS [ 5.430s] (1795/6817) moq-cli::integration import::import_delivers_the_catalog_finish_at_eof
PASS [ 0.839s] (1796/6817) moq-gst source::imp::session_tests::a_retired_rendition_drains_to_eos
PASS [ 1.196s] (1797/6817) moq-gst source::imp::session_tests::a_protocol_violation_posts_a_bus_error_after_playback
PASS [ 7.066s] (1798/6817) moq-audio encode::producer::tests::hls_stops_the_segment_where_retained_history_expired
PASS [ 2.367s] (1799/6817) moq-ffi test::shutdown_times_out_on_an_unfinished_track
PASS [ 2.157s] (1800/6817) moq-gst source::imp::session_tests::a_refused_redial_posts_a_bus_error_after_playback
PASS [ 2.287s] (1801/6817) moq-gst source::imp::session_tests::a_closed_publisher_session_resumes_on_the_next_start
PASS [ 7.061s] (1802/6817) moq-cli::integration report::import_reports_count_what_each_rendition_wrote
PASS [ 3.247s] (1803/6817) moq-gst source::imp::session_tests::a_restart_renders_through_a_synced_sink
PASS [ 5.871s] (1804/6817) moq-gst source::imp::session_tests::relay_loss_reconnects_after_refused_redials
PASS [ 10.277s] (1805/6817) moq-ffi test::session_state_names_a_failing_first_dial
────────────
Summary [ 15.550s] 1805/6817 tests run: 1804 passed, 1 failed, 114 skipped
FAIL [ 0.108s] (1782/6817) moq-gst source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad
stdout ───
running 1 test
test source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad ... FAILED
failures:
failures:
source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad
test result: FAILED. 0 passed; 1 failed; 0 ignored; 0 measured; 170 filtered out; finished in 0.07s
stderr ───
thread 'source::imp::session_tests::restarts_under_reconnects_interleave_on_one_pad' (1245554) panicked at rs/moq-gst/src/source/imp.rs:3788:13:
assertion `left == right` failed: [Pad("video_0"), Segment("video_0", 79000), Buffer("video_0", 1), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 94000), Buffer("video_0", 2), Buffer("video_0", 3), Buffer("video_0", 4), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 94000), Buffer("video_0", 2), Buffer("video_0", 3), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 62000), Buffer("video_0", 5), Buffer("video_0", 6), Buffer("video_0", 7), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 62000), Buffer("video_0", 5), Buffer("video_0", 6), Buffer("video_0", 7), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 73000), Buffer("video_0", 8), Buffer("video_0", 9), Buffer("video_0", 10), Start("video_0"), Caps("video_0", "video/x-h264"), Start("video_0"), Caps("video_0", "video/x-h264"), Segment("video_0", 2809000), Buffer("video_0", 11), Buffer("video_0", 12), Buffer("video_0", 13), Eos("video_0")]
left: 1
right: 0
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
warning: 5012/6817 tests were not run due to test failure (run with --no-fail-fast to run all tests, or run with --max-fail)
error: test run failed
mbx[cache]: object cache: 777 hits, 9 misses, 50 not looked up, 3 bypassed; 0 B downloaded, 0 B uploaded, 432.8 MiB stored locally
mbx[cache]: 1m 55s session; 27m 21s estimated compiler time avoided (summed across compilations)
mbx[cache]: 50 compilations had no usable prior inputs or matching prediction for a lookup
mbx[cache]: bypass reasons: 3 unsupported-crate-type
mbx[cache]: Cargo artifact reuse and CI cache archive transfers are not included above
mbx[savings]: 816567 compilations on file. the box remembers.
error: recipe `test` failed on line 294 with exit code 100
error: recipe `test-changed` failed on line 86 with exit code 100
##[endgroup]
##[error]ci: test rs failed
##[error]Process completed with exit code 100.
##[group]Run actions/upload-artifact@cf430e030ddbb5b0abf93d22962f4752f3646cd9
with:
name: cargo-timings
path: target/cargo-timings/cargo-timing-*.html
if-no-files-found: ignore
retention-days: 14
compression-level: 6
overwrite: false
include-hidden-files: false
archive: true
env:
BASE: 20955c09e94ae2dcc579060365f6dcee2855a328
CHECK_FAIL_FAST: 1
NEXTEST_TEST_THREADS: 24
MBX_CACHE_LINKS: 1
MBX_REMOTE_URL:
##[endgroup]
With the provided path, there will be 1 file uploaded
Artifact name is valid!
Root directory input is valid!
Uploading artifact: cargo-timings.zip
Beginning upload of artifact content to blob storage
Uploaded bytes 69996
Finished uploading artifact content to blob storage!
SHA256 digest of uploaded artifact is 7643f9635dbbd9f46e472050ecbccb000395e73b23a889e31d30374fc2ee0a16
Finalizing artifact upload
Artifact cargo-timings successfully finalized. Artifact ID 13174579057
Artifact cargo-timings has been successfully uploaded! Final size is 69996 bytes. Artifact ID is 13174579057
Artifact download URL: https://github.com/moq-dev/moq/actions/runs/41516535364/artifacts/13174579057
Post job cleanup.
##[start-action display=Run jdx/mr-boxington-action@47ce49b32025d93a7a2f4570815afaa2b7092240;id=__75c9f721-c833-4be0-b099-3bf6c7b94343.mbx]
##[end-action id=__75c9f721-c833-4be0-b099-3bf6c7b94343.mbx;outcome=skipped;conclusion=skipped;duration_ms=0]
Post job cleanup.
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git version
git version 2.55.0
Temporarily overriding HOME='/var/cache/github-runner-work/_temp/fcd029b8-caeb-4e74-b5c6-86a8307b0fa1' before making global git config changes
Adding repository directory to the temporary git global config as a safe directory
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git config --global --add safe.directory /var/cache/github-runner-work/moq/moq
Removing SSH command configuration
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git config --local --name-only --get-regexp core\.sshCommand
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'core\.sshCommand' && git config --local --unset-all 'core.sshCommand' || :"
Removing HTTP extra header
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git config --local --name-only --get-regexp http\.https\:\/\/github\.com\/\.extraheader
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git submodule foreach --recursive sh -c "git config --local --name-only --get-regexp 'http\.https\:\/\/github\.com\/\.extraheader' && git config --local --unset-all 'http.https://github.com/.extraheader' || :"
Removing includeIf entries pointing to credentials config files
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git config --local --name-only --get-regexp ^includeIf\.gitdir:
[command]/nix/store/lg46w7hrjx7kylsh6645c9l975i8h43b-git-2.55.0/bin/git submodule foreach --recursive git config --local --show-origin --name-only --get-regexp remote.origin.url
A job completed hook has been configured by the self-hosted runner administrator
##[group]Run '/nix/store/gx2v6q84b0w43bafm0whkdx93qyxbpps-ci-cache-gc.sh/bin/ci-cache-gc.sh'
shell: /nix/store/1mv3qz005gkbalxfghk4y3s8ayv4d2dl-bash-interactive-5.3p15/bin/bash --noprofile --norc -e -o pipefail {0}
##[endgroup]
evicted 263 objects and 49 action results (794.4 MiB logical); 49.9 GiB logical remain
dropped 3 stale checkout records
dropped 8 session event streams
removed 0 learned incremental units and 2 checkout directories (3.7 KiB logical); 51 units (4.7 GiB logical) remain
evicted 35 objects and 5 action results (508.3 MiB logical); 49.8 GiB logical remain
removed 2 target directories (13.7 MiB logical, 2 abandoned and 0 live); 12.7 GiB logical remain
Cleaning up orphan processes
(written by quest land watch) |
Contributor
|
codex / Codex land fixer note:
(written by codex / Codex) |
Contributor
|
quest land watch: re-enqueued this pull request at 4d1974a, at the back of the merge queue, unchanged, since the land fixer judged the failure flaky: (written by quest land watch) |
This was referenced Nov 3, 2026
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.
Problem
Below 2 Mbit/s, moq-quic can release two datagrams on one packet of pacing credit. Saturating subtraction discards the second packet's cost, allowing up to twice a controller's rate. This finishes
quest/m1/pacer-low-rate.mdand deletes it.Approach
Concrete source overlap checked before edits: moq-dev/moq#7602 owns build costs, currently fMP4 decoding and audit evidence. moq-dev/moq#7605 owns
bench/relay measurement support. Neither currently overlapsrs/moq-congestion/src/{pacer.rs,bbr3/mod.rs}. No status/FIN work duplicates moq-dev/moq#7439.Impact
offload_budget(send_quantumplus the delayed-ACK term) also drops by one datagram at low rates. It only binds when a connection negotiatesack_eliciting_threshold >= 4, so it has no effect under the default threshold (noted by the Pullfrog review).Measurements
Before is base
21d6da9fc108b48373a961f34e43c3fcd38b0146; after is this pacer correction plus the BBR3 quantum change. Every measured release run below enablesCARGO_PROFILE_RELEASE_DEBUG_ASSERTIONS=trueandCARGO_PROFILE_RELEASE_OVERFLOW_CHECKS=true. All expensive commands usequest gate.All 20 probe tests pass before and after. The table includes every paired comparison in that suite: the nightly scenarios use their complete seeds 1-8; short merge-gate comparisons use their complete seed set, seed 1. Values are means of per-seed outcomes, including per-seed p99. Probe overhead is redundant / unique media bytes.
All probe comparison rows
Larger competition sample: seeds 1-256, temporarily substituted only in
competing(); no budget or assertion changed, and the measurement override is absent from the final source. The command wasquest gate env CARGO_PROFILE_RELEASE_DEBUG_ASSERTIONS=true CARGO_PROFILE_RELEASE_OVERFLOW_CHECKS=true MOQ_PACER_SEEDS=256 just rs test --release -p moq-quic --lib --run-ignored only -E 'test(/^tests::probe::competing_/)' --no-capture, with that temporary seed override. Each 256-seed test passed before and after.Paired probing minus baseline, mean ± standard error over those 256 seeds:
The CUBIC late allowance is +4.5 frames; p99 allowance is +40 ms. Correcting pacing also changes the non-probing baseline, so the absolute late/p99 shifts must not be attributed entirely to probing. The quantum retune reduces the debt-only CUBIC paired spread rather than relying on a favorable eight-seed result.
Rerun after merging main through
30d6f63571moq-dev/moq#7458 (Quinn and BBR3 correctness sync) and GCC (moq-dev/moq#7546) landed after the measurements above. So the release probe suite, nightly cases included, was rerun on
origin/main(30d6f63571) and on this branch merged with it. The command wasquest gate env CARGO_PROFILE_RELEASE_DEBUG_ASSERTIONS=true CARGO_PROFILE_RELEASE_OVERFLOW_CHECKS=true RUST_LOG=info just rs test --release -p moq-quic --lib --run-ignored all -E 'test(/^tests::probe::/)' --no-capture --test-threads 1. All 22 tests pass on both. Same columns as above; main is the before.Every probe comparison row, main vs this branch, after #7458 and GCC
GCC's in-tree low-rate row is
rises_to_target_under_gcc: a 1 Mbit/s start on a 3 Mbit/s path that opens to 12 Mbit/s at 5 s. The correction leaves it unchanged. Late frames stay at 1 probing and baseline, and p99 at 73.76 / 72.76 ms. On every seed the target is reached 5 s (1x target) and 9 s (2x target) after the path opens. GCC'stests::realtimerows in #7546 came from an out-of-tree harness and can't be rerun here. BBR3 delay bound (moq-dev/moq#7544) and SCReAMv2 (moq-dev/moq#7547) are still unmerged. The realtime epic now records the GCC rerun and the need to use the corrected pacer for its bake-off.Validation
quest gate just rs test -p moq-congestion -p moq-quic --lib: 810 passed, 16 skipped, including Clippy and the final extended low-rate calibration test.quest check --base origin/main: 626 documents pass. Formatting andgit diff --checkpass.dc824374944af4cf2f557dbadf7c236f96d0514f(this branch merged with main as of then):quest gate nix develop --command just cipassed, exit 0, after deduplicatingPATHin the invoking shell with the same line moq-dev/moq#7594 adds to the dev shell'sshellHook. That's an environment fix, not part of this PR. No mbxE2BIGwas hit.just ciselects modules from the diff againstorigin/main: check rs, test rs (6526 passed, 110 skipped), markdown, gh, test gh, drill, quest (613 documents), shell, large_files, toml, nix, justfile. The flake check and JS suites weren't selected, so the host's known rootless-namespacecheck flakelimit and the flaky@moq/netovertaken-group test weren't reached.Argument list too long (os error 7)in Rustdoc and the Clippy-warning recipe fixture. One earlier attempt also hit a bind collision inbench/relay.test.py::PortOwnership.test_a_stream_transport_needs_its_own_listenerthat passed on rerun.96244dd854(moq-dev/moq#7458, which touchesrs/moq-congestion/src/bbr3/mod.rs) and then30d6f63571, needed no conflict resolution. Afterwardsgit diff origin/main...HEADmatches the reviewed diff except for hunk offsets and a quest-list context line. fix(quic): sync Quinn and BBR3 correctness fixes #7458 has nosend_quantum,smssfloor, or burst references, so it doesn't interact with the quantum retune.8b99e034b1:quest gate nix develop --command just cipassed, exit 0 (PATH deduplicated as above). Test rs: 6817 passed, 113 skipped. The first attempt, on5a9aeeb4da, failed onlymoq-tokio::integration backend::quic_client_close_drains_finished_trackwithsubscribe failed: Unroutableatrs/moq-tokio/tests/integration/backend.rs:902. That's a pre-existing flake: plainorigin/main(103e98e362) fails it 3 of 60 in--stress-count 60with the same panic. Branch heads failed it at similar rates: 5/30 merged, 3/30 pre-merge, 3/30 with only the pacer change, 3/30 with only the BBR3 change. One main run of 0/30 was luck.Alternatives
Keep saturating subtraction or compensate with larger pacing gains: both preserve incorrect rate accounting. Broad gain changes were unnecessary; the measured retune uses smaller feasible aggregates. No probe budgets or seed-count changes are included.
Choices made without asking
Used the existing BBR3 send-quantum mechanism to retune burst size rather than adding a knob or changing pacing gains. Used 256 paired seeds for the competing cases to separate their effect from the documented eight-seed noise. Kept all permanent seed counts unchanged.
Review
Head
4d1974a5ab74afe3457547fe12adf63f383bd5be: Codex,gpt-6.1-sol, medium effort, viaquest review --base origin/main(base24b04633a5), round 1: no findings. Since the previous reviewed headdc824374944af4cf2f557dbadf7c236f96d0514f, this head adds two conflict-free merges oforigin/mainand one quest-prose commit: the realtime epic records GCC's rerun row under the corrected pacer. Source is unchanged.Automated reviews on
dc824374944af4cf2f557dbadf7c236f96d0514f: the OpenAI review and Pullfrog reported no actionable findings. Pullfrog's note thatoffload_budgetalso drops by one datagram is now in Impact. CodeRabbit was rate-limited and didn't review.Earlier rounds on the source change:
Code review: Claude Code,
claude-opus-5-5, medium effort, viaquest review --base origin/main --harness claude.e8649b81a2347ac3a5dc1dca75d03af42f3c46f3: no findings.6b4dd0bfccb63a55c10bdb784f853852b05226c2, all four findings fixed:Follow-ups
Suggested: plan the pre-existing
moq-tokio::integration backend::quic_client_close_drains_finished_trackflake (subscribe failed: Unroutable, 3 of 60 onorigin/main) as a quest. Plan the observedbench/relay.test.py::PortOwnership.test_a_stream_transport_needs_its_own_listenerbind collision as a flake quest if it recurs. Benchmark A/A calibration still requires maintainer dispatch and was not claimed here.(Written by GPT-6.1 Sol; gate, quest completion, readiness, and the merge-and-rerun iteration by Claude Opus 5.5)