Skip to content

[Decision] Route CPU-heavy check:* runs through os-verify-lock? The measurement is in and one-directional — no cell of 64 reaches today's exit-99 baseline #13615

Description

@claude

[Decision] Route CPU-heavy check:* runs through os-verify-lock? — the number is in, and it is one-directional: no.

Provenance. Filed by the domain:devx seat (post #6023, session session_01Pk26oZ12t5N1hwGW1m1MgC), R33, carrying the measurement #12795 owed. Triage reserved this decision as a separate, probably-maintainer card and ⛔ this seat does not retire that reservation by having measured a clear answer. This is a confirmation, not a deliberation.

Measurement: #12795PR #13610, docs/audits/2026-08-verify-lock-gate-routing-measurement.md.

The number

Today 2.0% of locked runs return exit 99 (acquisition exceeded the 540s budget ⇒ the run is NOT MEASURED).

Across a 64-cell sweep — 4 routing policies × 4 demand levels × 4 levels of CPU credit granted to routing — the best cell is 4.0%, the worst 65.7%.No cell reaches today's baseline.

every routing policy measured makes the NOT-MEASURED rate worse, and "a routing rule that converts measurements into non-measurements is a net loss whatever it does to CPU" is the card's own criterion.

And it does nothing good to CPU either: W=1 was never the throughput maximum in 3 of 3 concurrency sweeps — unsaturated box W=4 beat W=1 by 2.38×, saturated box flat.

The model was validated before it was trusted: replayed against the population --report prints, simulated exit-99 2 vs observed 2, wait p50 151s vs 151s, p90 536s vs 536s. The other 63 cells are evidence because the calibration cell reproduces reality on three independent statistics.

四维分析

1 — 实际业务需求。 今天没有人因为门禁不走锁而受损:基线 exit-99 是 2.0%,而任何一种路由都把它抬到 4.0% 起。⇒ 采纳路由是用一个真实的可用性退步,换一个纸面上的"名副其实"。⛔ 没有业务需求推动它。

2 — 平台长远合理性。 锁的名字与它的保证确实不一致(#12528 已把这条边界写进文本)。但一致性的正确修法是改名字或改文本,不是改行为 —— 尤其当改行为被实测证明会让它服务的那个量变差。⭐ 而且还有一条预算硬事实:最重的两个族合计 534.5s,而本文件的预算是 540s一次被路由的重族清扫,就能吃掉另一个调用者的全部获取预算。

3 — 避免 AI 写代码犯错。 ⭐ 本卡最该记的一条:本席在派发令里写了一条错的前提 —— 「#12538 / PR #12788 把等待预算抬高了」。实测 HARD_CAP_S=540DEFAULT_WAIT_S=540 都没动,动的是 SLOT_MAX_AGE_S(→1620s)。⇒ 那条围栏朝宽松方向错,会让路由显得比实际更负担得起。⚠️ 一条错向宽松的围栏,正好招来它本该拦住的那个改动。

4 — 创业阶段不扩散需求。 采纳「什么都不路由」的成本是:不改代码、不改预算、不改锁的行为。#13610 已经把这条结论写进锁的头注(退掉那句现在为假的「此权衡未经测量」),⇒ 裁定「维持现状」不需要任何后续工程动作。

三条候选

  • A —— 维持现状,什么都不路由。(本席建议)零成本;实测支持;锁的头注已记录测量与结论。
  • B —— 路由全部 CPU 重族。 ⛔ 被实测否决:64 格无一格达到基线。
  • C —— 只路由最重的 5 个族(占门禁总秒数 64.3%,前 2 个占 46.3%)。 ⚠️ 分布确实够尖锐,能划出干净阈值 —— 但它仍然让 Q3 变差,因为最重的两族合计 534.5s 对 540s 预算。

⚠️ 唯一数字settle不了的子问题(来自 dev 的 option C,原样上呈)

路由那 5 个重族,是用测量的「可得性」去买测量的「卫生」。

即:让重族串行,能让它们彼此之间的读数更干净(不互相抢核),代价是抬高别人拿不到锁的概率。⭐ 这是一个价值判断,不是一个测量问题 —— 本席不裁,原样交上。

⛔ 本席的边界

⛔ 不代裁:路由决策触舰队形态/支出,按人工地板归维护者;triage 也明确把它留成独立卡。上面是建议与数字,不是裁决。

Refs: #12795(测量卡)· PR #13610(审计文档 + 锁头注更正)· #12528(披露半边)· #13611(顺带发现:轻量族 ~97% 的时间是 pnpm 启动开销)· #12783(ledger 位置,限制了 --report 能回答什么)


Generated by Claude Code

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions