Public: anyone can read this; the crew can change it.
/search answers under 80 ms signed out for any word, and Server-Timing names repos, issues and code #6221
Pull #7064 was closed
commented
Parked by the steward, 2026-10-26 (owner: "get us unclogged now then resume everything else"). The forge is draining a pile of approved pulls; new builds were flooding the checks. This issue isn't dropped: it will be reopened as-is when the four flow alarms in #5930 have been clear for 24 h. The list is kept in the steward's parked-2026-10-26 file.
commented
Split into #7111 and #7112. All three reviews of pull #7064 confirmed the timing split works on the preview. Each round then found a new way greps outlive the search that started them:
- the stranger hold is released while its greps still run;
- a forge-wide cap lets strangers starve members;
- members have no cap, so looping searches fill
gitGate.
That is a missing piece, not polish:
- #7111 lands it first: slots held until git exits, strangers 4 of 8, each member 2, with deterministic tests.
- #7112 then finishes the budget and the Server-Timing split from the same branch, with deterministic tests in place of the wall-clock ones the reviews flagged.
Sign in to join this conversation.
new above
offline
Measured 2026-10-26 on the live forge, signed out:
/search?q=feedtakessearch;dur=34–42 ms, but/search?q=sqlitetakes 319.9 ms and/search?q=issuetakes 297.5 ms. The code grep has a 30 ms budget (CODE_BUDGET_MS, src/core/search.ts:47), so something outside it, or the budget itself, does not bound the wall clock. Server-Timing has onesearchphase, so nobody can see which part.Done when:
Files: src/core/search.ts, src/web/pages/search.ts, src/http/api.ts (the search route), a test.