docs(code-index): serving-current latency after seat keep - #1567
Closed
ScriptedAlchemy wants to merge 4 commits into
Closed
ScriptedAlchemy wants to merge 4 commits into
ScriptedAlchemy wants to merge 4 commits into
Conversation
A retryable graph activation used to erase the prepared serving candidate, and an unfinished clone-fingerprint successor withheld the same seat after exact and lexical owners were ready. Keep the candidate in both cases so search can move off the predecessor while graph retries. Co-authored-by: Zack Jackson <ScriptedAlchemy@users.noreply.github.com>
Co-authored-by: Zack Jackson <ScriptedAlchemy@users.noreply.github.com>
|
Co-authored-by: Zack Jackson <ScriptedAlchemy@users.noreply.github.com>
Owner
Author
Cross-linkAgrees with keeping Also see:
Flag: if status-census wakes requeue reconcile for ~30–60s after proof expiry, land that cut even if #1568 removes successor from the receipt — otherwise search-can-hit / status-not-fresh can still strand the helper. No merge from this comment. |
This was referenced Sep 17, 2026
This was referenced Sep 18, 2026
Contributor
Performance Comparison
|
Owner
Author
|
Closing as stale. Evidence from the triage: git diff --stat origin/master...origin/cursor/graph-rebuild-current-latency-fb0d = docs/code-index-serving-current-latency.md +83 PLUS the 3 files of b80dd58; doc-only once #1562's fix is attributed to #1562; merge-base 918 behind Reopen with a rebase if the remaining value is wanted. |
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
2 participants
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
Investigation only. Draft. Do not merge. Does not change the 90s receipt timeout.
Stacks on #1562 (
b80dd58). The new commits are the note indocs/code-index-serving-current-latency.md.Local receipt: after the seat-candidate keep, the serving seat matches and search can hit inside ~90s, but
wait_for_current_generationstill hitsRECEIPT_TIMEOUT. Timeout status hasseated_generation_age_seconds≈ 89. A mid-run sample puts the test thread in that helper atgraph_rebuild_status_test.rs:174(thetimeout(...).awaitaround the loop at148-172), inside status/tool MCP calls. The prior fail did ~16k status polls. Search is not that park: it runs only after status already looks current (153-160).Age is seal time (
status.rs:293-296), so the replacement sealed about a second into the wait. Estimates are structural, not a profiled re-run.Wait helper vs production current
The helper (
graph_rebuild_status_test.rs:140-179) matches productioncurrent(status.rs:451-489): coveragecomplete, stalenessfresh, graphReady, seat match (registry.rs:1205-1231). Search only needs the seated exact/lexical owners. Do not treat seat match plus hits ascurrent, and do not raiseRECEIPT_TIMEOUT(graph_rebuild_status_test.rs:26). Do not add a sleep to thin the polls.Status poll hits capture and project
tracedecay_statusalways snapshots the generation census (crates/tracedecay/src/mcp/tools/handlers/dispatch_groups.rs:533-541, read atstatus.rs:257) even though this test already disables branch diagnostics, storage health, session ingest, and staleness (graph_rebuild_status_test.rs:103-109). The census reader isProjectCodeGraphServingAuthorityV1::project(project_reads.rs:86-139).projecttries the ready-decoded seat first (serving_reads.rs:965-988). That callsready_without_stat(serving_reads.rs:928) →GitMetadataFingerprintV1::capture(reconcile.rs:540-561,identity.rs:167-176) and, on a miss,note_wake_if_idle(serving_reads.rs:924-930). The next arm,retained_text_owner_freshness_for_scope, captures again (serving_reads.rs:1291-1308). Only then does it use the O(1) seat (serving_reads.rs:1342-1364), which neither captures nor wakes.captureis a few git-metadata stats (identity.rs:167-176). ~16k of them are about 1–5s of filesystem, not the 89s. The activation coupling is the wake.note_wake_if_idlecoalesces (registry.rs:2133-2151), but once a pass drains the arrival the next poll posts another. The proof expires at 30s (code_index_scheduler.rs:36, checked atreconcile.rs:562). After that, an unchanged seat still failsready_without_stat, and the census keeps requeueing the worker intoReconcilePassGuard(mount.rs:539). Freshness stays offfreshwhile search can already hit. The freshness ladder the helper reads does not callcapture(serving_reads.rs:324-333).Post-seal serialization
Published pass, one worker:
mount.rs:1109-1141, stop atregistry.rs:2330-2332). Clone pages are not in that await.serving.rs:3204-3217and2650-2658) after dropping the text reservation (serving.rs:3188). Graph then runs under the 128 MiB clone reservation. Full overlap with the text reservation is the RSS failure atmount.rs:1052-1057.mount.rs:1463-1490).activateis awaited before the swap (mount.rs:1601-1706, swap at1818). Retryable failure keeps the candidate (1676-1681). Resident-memoryBudgetExhaustedis a refusal (reconcile.rs:360-367), sodashboard_generation_is_readystays false.mount.rs:1747-1782). The swap already bindspass_proves_latest(1862-1867).mount.rs:1735-1740) contradicts the post-swap idle rule (1978-1991) and re-enters the pass guard before source reconcile (539-541, drop at1172-1174).load_active_sharedunder the scheduler mutex and the serving write lock (mount.rs:1854-1870,ignored_dependencies.rs:158-165).This test compiles the 50ms activation floor (
registry.rs:87-90), not the production 30s.Proposals
#1562-eligible
reconcile_in_progressacross source reconcile for a successor-only continuation (mount.rs:539-541vs1172-1174). About 2–15s of post-seat non-current. Not the whole 89s.mount.rs:1735-1740), but yield once after the swap with the pass guard down. Fixing only this wake is not enough: the status census posts the same wake (serving_reads.rs:924-930and1296-1308).Wait helper
project(project_reads.rs:86-96) should take graph statistics from the seated generation (serving_reads.rs:1342-1364) and must not callready_without_stat,GitMetadataFingerprintV1::capture, ornote_wake_if_idle. Query admission can keep the wake. Up to ~30–60s ofverifyingafter the 30s proof expiry (code_index_scheduler.rs:36), which is the activation stall the sample explains. The 1–5s of fingerprint syscalls are not the 89s by themselves.dispatch_groups.rs:533-541). This test already turns the other sections off (graph_rebuild_status_test.rs:103-109) and still paysproject. The helper does not readgraph_statistics. Combined with (3), the test thread stops parking incaptureandprojectatgraph_rebuild_status_test.rs:174.Follow-up
begin_clone_successorin the owner-installing advance (serving.rs:3212-3217and2654-2658). If the dump shows a resident-memory refusal, this is never-currentversus one publish. Otherwise a few seconds.mount.rs:1754-1782). About 1–8s.builder.rs:32-70) with text advances. Keepclaim_build(code_graph.rs:1244-1316) afterserving.rs:3188. About 5–20s. Do not fully overlap without an RSS measurement.load_active_sharedonly on a cache miss. About 0, or 5–20s on a miss.Do not change
captureandproject, andprojectstill wakes the worker.current.Checklist