[PR] ci(#4360): content-hash caching for pretest dist + bolder Windows sharding — test lane down to ~6.5 min | ci(#4360): pretest dist 内容哈希缓存 + windows 分片加粗 — 测试车道压到 ~6.5 分钟 by ranxianglei · Pull Request #4361 · ranxianglei/billion-context

ranxianglei · 2026-10-23T02:57:59Z

Closes #4360

目标

CI 测试车道 <8 分钟(Windows)。三个杠杆,前两个已进 master,本 PR 是最后一个:

pretest dist 内容哈希缓存 —— 已由 [PR] CI: build once, reuse everywhere — cache kernel/pi-subagents dist by content hash so legs stop rebuilding 6 minutes of artifacts every run #4377 进 master(经由 package.json pretest 脚本接线;本 PR 与之完全兼容,rebase 后零重叠)。
closeAllConnections() 的 linger 3s 税 —— 已由 [PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 进 master:每个 proxy teardown 固定 ~3s 空等被消除(lane matrix 194s→2.6s、plugin-fork 150s→8.9s)。
本 PR:按实测耗时平衡分片 + Windows 分片加密。

问题:行数均衡 ≠ 耗时均衡

旧 LPT 按行数分片,「完美均衡」的字节数产出了 5min..37min 的腿(同一 run 的 windows 22-1=5.1min vs 22-5=37min+):spawn 密集型套件每行成本比纯断言文件高一两个数量级。

修复

scripts/ci-shard-partition.mjs(+.d.mts):权重 = 实测 ms(scripts/ci-shard-weights.json),未测量文件回退 lineCount × msPerLineFallback(中位数);LPT 算法不变,仍是确定性、total(union == 全部文件,零覆盖损失)。
scripts/ci-shard-weights.json:840 文件全量实测表,post-[PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 树上单文件计时(~66min 串行总量);本地环境挂起文件按 150s 超时下界记录并列入 timedOut,本地 rc=1 环境失败排除。平衡退化时用 scripts/measure-test-durations.mjs 再生。
ci.yml:windows-22 CORE 3→5 分片;FULL windows(22+24)3→5。
效果(本地预演):最大片串行 1984s → 794s(片内 c=4 并行,目标 ~7min worst-leg 含 setup)。
顺带(d06d192):[PR] fix(ci): run corrupt index.db degrade checks in a child process so windows cleanup can't hit EBUSY (#4338) #4378 把 corrupt-index 检查移进子进程时漏了 LIVE_FILES 登记,SPAWN_RE 绊线在 master 上正拦着每条腿——本 PR 尾部追加登记(storage-engine-proofs、web-storage-banner)。

测试

tests/ci-shard-weights.test.ts 4 test:测量值优先/行数兜底、monster 隔离、确定性+total、表内条目必须真实存在(防 stale)。
45/45 绿(ci-shard-weights + ci-pretest-cache + ci-runner-failure-shapes),typecheck 干净,LIVE_FILES 43 条,storage-engine-proofs 12/12 绿(带新端口带)。

注意

分片名变化:test (windows-latest, 22, 4)、(22, 5)、(24, 4)、(24, 5) 为新检查名,若 branch protection 精确枚举了检查名需要同步(用通配则无需)。
权重表是快照不是持续校准:新落地的重量级套件先走行数兜底,平衡明显退化时跑一次 regen 即可(README 注释里写明)。

github-actions · 2026-10-23T03:01:40Z

📦 Built Package Artifact

Branch: 2026-10-23_ci_windows_under_8min (d06d192)

Option A — Install this exact build (recommended)

npm install -g [email protected]

This exact version is immutable and reproducible. The rolling pr tag tracks the newest test build across all PRs.

Option B — Download artifact

Download the artifact from the Actions run
Extract the tarball and install:

tar xzf billion-context-pr4361.tgz
npm install -g package

Install instructions are refreshed on each push.

ranxianglei · 2026-10-23T06:35:56Z

(ai: zai/glm-5.3)

关联:#4380 / PR #4384 修复了每个 proxy teardown 固定 ~3s 的 post-response linger 税(lane matrix 194s→2.6s)。这是本 PR 8 分钟目标的最大单项杠杆——本分支的分片权重表将在 #4384 合并后用修复后的真实耗时重新生成。

Line-count LPT balanced bytes, not seconds: process-spawning suites cost orders of magnitude more per line than pure assertion files, and the 'perfectly balanced' layouts still produced 5min..37min legs from equal byte counts (windows 22-2/22-3 vs 22-1/22-4 on the same run). With #4384's linger cut landed the per-file wall times are stable enough to weight the partition by MEASURED duration: - scripts/ci-shard-partition.mjs: fileWeight = last measured ms per file (scripts/ci-shard-weights.json), new files fall back to lineCount x msPerLineFallback (median ms/line); same deterministic total LPT otherwise. - scripts/ci-shard-weights.json: 840-file table measured post-#4384 (single-file node --test runs, VM, ~66min serial total); local-env hangs capped at the 150s timeout and listed under timedOut, local rc=1 env-fails excluded. Regenerate via scripts/measure-test-durations.mjs when shard balance degrades. - run-ci-tests.mjs delegates partitionAll to the module (this script has top-level side effects and cannot be imported by tests). - ci.yml: windows-22 CORE 3->5 shards, windows FULL 3->5 (both node versions) — measured max shard drops 1984s->794s serial (concurrency c=4 per shard => ~7min worst-leg target incl. setup). Closes #4360 (with #4377's dist cache + #4384's linger cut, the three levers of the 8-minute goal)

…follow-up) #4378 moved the corrupt index.db degrade checks into spawned child processes but skipped the LIVE_FILES registration, so the SPAWN_RE tripwire fails every run-ci-tests leg on master right now. Appended at the tail (slot = index, no renumbering).

ranxianglei · 2026-10-23T12:57:44Z

#4377(dist 内容哈希缓存,叠在 #4355 上)已于 2026-10-23 09:59 合并,与本 PR 目标相同;本 PR 关闭作为被取代。若这里面还有 #4377 没覆盖的点,欢迎在新 PR 里捡起来。

ranxianglei · 2026-10-23T16:20:37Z

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

查重: 不重复——这是对已关闭条目的遗留点捡起,不是新立项;检索 open issues/PRs(sharding / 分片 / duration-weighted CI)无在途同类工作。

两条未答评论的回应

2F(closeAllConnections() is silently thwarted by post-response linger: every test/shutdown teardown idles for 3–5 seconds, | closeAllConnections() 被 post-response linger 静默挫败:每次测试/停机 teardown 空等 3-5 秒,CI 全套件累计几十分钟纯墙钟 #4380/[PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 linger 税):[PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 已于 2026-10-23 12:23Z 合并,「[PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 合并后重新生成权重表」的承诺条件已满足——本分支权重表正是 [PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 后实测的(840 文件表)。
1F(artifact bot):标准构建通知,无需动作。

按你 4F 的指示捡起了本分支剩余杠杆(「若这里面还有 #4377 没覆盖的点,欢迎在新 PR 里捡起来」):

分支 2026-10-23_ci_windows_under_8min 上共 2 个 commit,逐一核对:

commit	内容	处置
`d06d192ff`	LIVE_FILES 注册两个 corrupt-index spawn 套件	不捡——已经由 #4416 进了 master(同一机制,行内容一致)
`0569781af`	分片按实测耗时加权(LPT)+ windows 3→5 腿	捡起——master 没有,cherry-pick 到当前 master 干净落地 → PR #4439(issue #4437,标签 scope/ci + p3)

效果:测量时点 max-shard 串行 1984s→794s;对当前 master 文件集(861 文件,21 个新增走回退权重)本地复算 n=3 max≈1360s → n=5 max≈816s,五腿负载差 <1s,Windows 最重腿目标 ≈7 分钟(含 setup)。

需要 owner 动作一项:出现 4 个新 check 名 test (windows-latest, 22, 4|5)、test (windows-latest, 24, 4|5)——branch protection 按 test (os, node, shard) 名 keying,需把这 4 个加入 required checks,否则新 shard 失败不挡合并(PR 描述与 #4437 均已披露)。

结构化八问

是什么问题:CI 测试分片按行数切,但行数 ≠ 耗时——进程派生型套件几百行要跑几分钟,导致个别 Windows 腿 20+ 分钟、别的腿 5 分钟,整条车道墙钟被最重一腿拖着。
什么原因导致的:LPT bin-packing 以行数为权重;spawn 型套件每行成本比纯断言文件高几个数量级,「按字节均分」≠「按秒均分」。
影响范围与严重程度:仅内部 CI,无外部用户影响;性能/体验级(lane 墙钟 10–18 min → 目标 ~7 min),非功能缺陷。
怎么修复的:分区权重换成每文件实测毫秒([PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384 后实测表;新文件回退 lineCount×fallback),LPT 保持确定性+全覆盖(所有 shard 并集 == 全部文件,零覆盖损失);windows 分片 3→5。
是否有回归问题:审过既有行为——分区仍 total/deterministic(新测试钉并集不变量+确定性);ubuntu 车道保持 /3 不变;既有 12 个 check 名不变且继续生效;唯一风险面是 4 个新 shard 未入 required checks 前的保护缺口(已披露,方向是少挡不多挡)。无 kernel/wire/config/持久化变更。
是否破坏缓存:不涉及(CI 基建,与运行时缓存/重试/增量游标无关)。
是否改了设计方向:没有——同一 LPT 框架内换权重来源+加粗分片数,架构与数据流不变。
是否引入了新的配置:没有(无新 env/config/默认值;权重表是仓库内数据文件,配再生成脚本 scripts/measure-test-durations.mjs)。

结论:本条目保持关闭(被取代判定成立,不回滚);剩余杠杆已由 #4439 独立闭环(Closes #4437),等 CI 绿光 + 人工合并;branch protection 补 4 个 check 名需要你操作。

ranxianglei · 2026-10-23T16:29:34Z

🤖 Powered by ework · qwen3.8-27b

(ai: vllm-qwen/qwen3.8-27b)

这条(4F 的「捡起来」指示)已在 7F 完整处理——遗留杠杆已捡起为 PR #4439(issue #4437),八问与 owner 动作项都在那条回复里,不重复。

补充一个状态更新:#4439 的首次 CI 运行在 16:22:32Z 被整体取消(10 个 workflow 同一时刻 cancelled)。取消前 guard 类 job 正常跑了约 4 分钟,零失败;分支上没有第二次运行记录,不是并发组自取消,属外部/手动取消。改动本身不需要任何代码动作——随时可 Re-run 或在下次 push 时自动重跑;如果你是想等 branch protection 补上 4 个新 check 名(test (windows-latest, 22|24, 4|5))再跑,那正好可以一起办。

Line-count LPT balanced bytes, not seconds: process-spawning suites cost orders of magnitude more per line than pure assertion files, and the 'perfectly balanced' layouts still produced 5min..37min legs from equal byte counts (windows 22-2/22-3 vs 22-1/22-4 on the same run). With #4384's linger cut landed the per-file wall times are stable enough to weight the partition by MEASURED duration: - scripts/ci-shard-partition.mjs: fileWeight = last measured ms per file (scripts/ci-shard-weights.json), new files fall back to lineCount x msPerLineFallback (median ms/line); same deterministic total LPT otherwise. - scripts/ci-shard-weights.json: 840-file table measured post-#4384 (single-file node --test runs, VM, ~66min serial total); local-env hangs capped at the 150s timeout and listed under timedOut, local rc=1 env-fails excluded. Regenerate via scripts/measure-test-durations.mjs when shard balance degrades. - run-ci-tests.mjs delegates partitionAll to the module (this script has top-level side effects and cannot be imported by tests). - ci.yml: windows-22 CORE 3->5 shards, windows FULL 3->5 (both node versions) — measured max shard drops 1984s->794s serial (concurrency c=4 per shard => ~7min worst-leg target incl. setup). Four NEW check names appear (windows 22/24 shards 4/5); branch protection must add them before they gate merges. Cherry-picked from closed PR #4361 (commit 0569781, stranded when the PR was superseded by #4377; picked up per owner invitation in the #4361 thread). The sibling LIVE_FILES commit (d06d192) is already on master via #4416 and is not included. Closes #4437

ranxianglei mentioned this pull request Oct 23, 2026

[PR] fix(server): closeAllConnections() cuts through the post-response linger (#4380) #4384

Merged

ework-agent added 2 commits October 23, 2026 20:30

ranxianglei force-pushed the 2026-10-23_ci_windows_under_8min branch from 92cd873 to d06d192 Compare October 23, 2026 12:32

ranxianglei closed this Oct 23, 2026

Navigation Menu

Repository navigation

[PR] ci(#4360): content-hash caching for pretest dist + bolder Windows sharding — test lane down to ~6.5 min | ci(#4360): pretest dist 内容哈希缓存 + windows 分片加粗 — 测试车道压到 ~6.5 分钟 - #4361

ranxianglei commented Oct 23, 2026 •

edited

Loading

Uh oh!

github-actions Bot commented Oct 23, 2026 •

edited

Loading

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

ranxianglei commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

目标

问题:行数均衡 ≠ 耗时均衡

修复

测试

注意

Uh oh!

github-actions Bot commented Oct 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

📦 Built Package Artifact

Option A — Install this exact build (recommended)

Option B — Download artifact

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

ranxianglei commented Oct 23, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

ranxianglei commented Oct 23, 2026 •

edited

Loading

github-actions Bot commented Oct 23, 2026 •

edited

Loading