Skip to content

[PR] fix(dsh-drift): name the oversized top-level field in the egress-fuse breakdown + document the market-disable brick (#6957 #6956 #6968 #6972) - #6984

Open
ranxianglei wants to merge 6 commits into
masterfrom
2026-11-17_dsh-drift-batch
Open

ranxianglei wants to merge 6 commits into
masterfrom
2026-11-17_dsh-drift-batch

Conversation

@ranxianglei

@ranxianglei ranxianglei commented Nov 16, 2026 •

Copy link
Copy Markdown
Owner

Summary

Root-cause batch for the dsh-drift aggregation (e-work #6983 / GitHub #6980): the dsh lane does not survive host-side state changes (plugin disabled, supplier switched, panel fallback, uninstall). This PR lands the in-repo fixes and documents the DSH-side asks.

Per-issue breakdown

Behavior change

  • The egress-fuse / doom messages now include an (other: <field> ~N, …) suffix when the other bucket carries real mass. This is operator-facing (the dsh UI error), not model-facing. No wire, config, or persistence change. The total/messages/tools/system/other breakdown values are unchanged (additive otherTop field).

Verification

  • tests/side-request-isolation.test.ts: new otherTop + measuredBreakdownClause tests (both directions: field named when mass present, no suffix when absent).
  • tests/raw-body-walk-depth.test.ts: otherTop: [] added to the deepEqual expectation.
  • tests/issue4079-egress-window-fuse.test.ts + tests/issue4185-identified-fresh-forward.test.ts: pass (regex-anchored doom-message assertions unaffected).
  • tsc --noEmit clean.
  • CI: no NEW failures — the architecture-docs + 3 test shards (ubuntu 24,3 / ubuntu 22,3 / windows 22,4) are red on master identically (pre-existing).

Related #6956 #6957 #6968 #6972

ci-check and others added 5 commits November 16, 2026 20:42
The probe anchor PROXY_TOOL_AT=6 was calibrated on 2026-10-31 for #4521
(schedule {4,6,8,9}). The nudge scheduling has shifted again — nudge now
lands at {3,5,11,14}. Re-calibrate the anchor to the next nudge-carrying
request pair (5 → tool round, 6 → fold round). The fold/proxy-tool-round
mechanics under test are unchanged.

Closes: master-red on test (windows-latest, 22, 3) and (ubuntu-latest, 22, 1).
The probe's hardcoded request-index anchors (PROXY_TOOL_AT/FOLD_AT) broke
every time the nudge gate's schedule shifted. The last recalibration
(245ecb1) was measured on a direct local run (timestamps ON by default
since #4313), but CI runs the suite preload pin BILI_MESSAGE_TIMESTAMPS="0"
(timestamps OFF) — a different wire shape, a different schedule (no nudge at
index 5) — so it passed locally and failed on every CI shard.

Root-cause fix instead of another anchor whack-a-mole:
- Arm the fold on the first request carrying the ephemeral nudge with >=11
  tagged refs; fold on the re-request after the acp_status round. The probe
  now materializes the #2929 scenario wherever the schedule places it.
- Pin compress.messageTimestamps:false so the probe's wire shape is hermetic
  regardless of how the suite is launched (matches the CI preload pin).

Supersedes the 245ecb1 recalibration. Verified: passes with and without
the preload pin, same deterministic schedule (nudge at inputs[7], fold at
inputs[8]).
…6957)

The #554 guard meter splits the outbound body into messages/tools/system/other,
but the 'other' bucket was an opaque aggregate — a 1.6M-token 'other' (#6957,
the dsh→official-DeepSeek wire put the mass in a top-level field bili's
messages-only compression can never touch) told the operator a dimension was
oversized without saying WHICH field, so the doom message's 'trim the oversized
dimension' was unactionable and the session sat in an egress-fuse dead loop.

estimateRawBodyBreakdown now tracks the top-level field(s) carrying the 'other'
mass (otherTop, sorted desc, exhaustive over the bucket) and
measuredBreakdownClause names the top-3 in the fuse/doom messages when 'other'
carries real mass. Additive: total/messages/tools/system/other are unchanged,
so the existing bucket tests and the regex-anchored doom-message tests keep
passing.

Also documents the #6956 market-disable path in dsh.bundle.patch.yml: the dsh
plugin-market 'disable' toggle kills BOTH entries (bili-native AND the gate) at
once, so the gate service never registers, the presets' inject latch is
unsatisfiable, and the host reports preset-ptc as not-found, bricking
new-conversation init. That is a DSH-side defect (an unsatisfiable inject must
degrade, not hang); the supported off-switches are disabling bili-native only
(#3662) or removing billion-context from bundles (#6972).

Closes #6957
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

📦 Built Package Artifact

Branch: 2026-11-17_dsh-drift-batch (02948bb)

Latest build — install this exact version

npm install -g billion-context@0.0.10502-pr.6984.1bddb3f

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

Recent builds (commit → version)

commit version build
02948bb billion-context@0.0.10502-pr.6984.1bddb3f build
117540a billion-context@0.0.10498-pr.6984.c5b1a66 build

Download artifact

  1. Download the artifact from the Actions run
  2. Extract the tarball and install:
tar xzf billion-context-pr6984.tgz
npm install -g package

Install instructions are refreshed on each push.

@ranxianglei ranxianglei added bug Something isn't working dsh p2 root-cause labels Nov 16, 2026
@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷
(ai: vllm-qwen/qwen3.8-27b)

查重: 不重复 — 镜像 open [PR] 行中仅本条指向 dsh-drift 聚合,无并列 PR;上游 PR #6984 即本线程唯一修复 PR。
标签: bug, dsh, scope/wiring, p2, pr(出站 egress-fuse 守卫属协议/出站边界域;诊断性修复,能用但有明显问题 → p2,与 PR 自身标签一致)。

Review of PR #6984 (head 2026-11-17_dsh-drift-batch)

Diff scope (5 files, +88/−10): src/preflight.ts (otherTop tracking in estimateRawBodyBreakdown), src/server.ts (measuredBreakdownClause suffix), dsh.bundle.patch.yml (#6956 comment note), 3 test files. No wire/config/persistence/kernel change.

Code review — no blockers:

CI (verified live): previous run red on architecture-docs + #2929 test shards — both inherited master reds (master itself fails architecture-docs: reference/architecture.md missing: src/responses-system-sections.ts, introduced by #6922; #2929 root cause fixed on ci-fixes 724741097; convergence PR #6974 still awaits your merge). Per the CI coordination policy I merged origin/ci-fixes into the PR branch and pushed — new head 02948bb15f (clean merge, zero conflicts; the 5 PR-touched files had no master-side churn). Posted the one-line report to #4599. Local verification on the new head: architecture-docs guard green, targeted 57/57 (side-request-isolation, raw-body-walk-depth, issue4079, issue4185, issue2929), bundle-patch tests 16/16, tsc --noEmit clean. CI re-running on the new head — I'll confirm when it lands.

结构化八问

  1. 是什么问题:dsh→官方 DeepSeek 线路把 ~1.6M token 塞进一个顶层字段,bili 的 messages-only 压缩碰不到它,出站熔断 doom 消息只说 "other" 超量、不说哪个字段,操作员无法行动,会话卡在熔断死循环。
  2. 什么原因:side request(max_tokens<=200)绕过 preflight 且不检查载荷体积,携带全量历史直转上游导致 400 #554 守卫计量把出站体拆成 messages/tools/system/other 四桶,但 other 桶是不透明聚合——walk 计 token 时丢弃了顶层字段名,诊断信息先天缺"哪个字段"这一维。
  3. 影响范围与严重程度:仅影响熔断/doom 触发时的报错文案(操作员可见,dsh UI);触发条件 = 出站体超窗且质量在 other 桶。本 PR 不改熔断行为本身;DSH 侧的巨型字段(bricking 根因)不在本 PR 范围,由聚合单跟踪。
  4. 怎么修复的:walk 在 other 桶计数时同步记录顶层字段(otherTop,降序、对桶穷尽),measuredBreakdownClause 在 other 有真实质量时把 top-3 字段名附进熔断/doom 消息;[Bug] Unable to create a new conversation after disabling the billion-context plugin (suspected to be caused by the Patc | [Bug] 关闭 billion-context 插件后,无法新建对话 (疑似 Patch/Gate 机制未回滚导致) #6956 市场禁用 brick 以注释形式记入 dsh.bundle.patch.yml 并指向 docs/clients.md 的受支持关闭方式。
  5. 是否有回归:无。四桶数值逐字节不变(纯增量字段);3 处消息调用点均为操作员侧错误响应,不进模型上下文;regex 锚定的 doom 断言与 deepEqual 钉扎均过;bundle 注释对 YAML 惰性、3 个 bundle 测试 16/16 绿。
  6. 是否破坏缓存:不涉及(纯诊断计量,不触碰缓存/重试/游标)。
  7. 是否改了设计方向:不涉及(守卫计量内部增量,边界不变)。
  8. 是否引入新配置:不涉及(无新 env/config/默认值)。

行为变更披露(旧→新):熔断/doom 消息在 other 桶有质量时尾部新增 (other: <field> ~N, …) 后缀——旧 …· other ~1,600,000 → 新 …· other ~1,600,000 (other: <field> ~1,600,000);仅操作员可见,PR 正文已披露。

判定:代码无阻塞项,可合——等 02948bb15f 的 CI 重跑转绿后由 owner 合并;#2929 与 architecture-docs 的继承红已随 ci-fixes 并入消除,最终收敛仍待你合 #6974(ci-fixes→master)。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷
(ai: vllm-qwen/qwen3.8-27b)

CI 状态更新(head 02948bb15f):29/30 完成、0 failed —— 10/10 test 分片全绿(含此前继承红的 #2929,ci-fixes 根治已生效)、architecture-docs 绿、static-gates / mysql-contract / 全部 guard 绿。仅剩 test (ubuntu-latest, 24, 3) 一个分片仍在跑(约 50 分钟,该分片正是 #2929 历史红分片,大概率是 runner 负载慢而非失败)。转绿我补一行确认;若红,拉日志定向排查。

@ranxianglei

Copy link
Copy Markdown
Owner Author

🤖 Powered by ework · qwen3.8-27b

[bot] 🏷
(ai: vllm-qwen/qwen3.8-27b)

CI 全绿确认(head 02948bb15f):30/30 完成 —— 14 success / 16 skipped(e2e 车道按 env 门设计跳过)/ 0 failed;最后慢的 test (ubuntu-latest, 24, 3) 也已 success。审查结论不变:代码无阻塞项,CI 绿,可合——等 owner 合并 PR #6984。

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant