Long RPCs ignore cancellation mid-operation; a done context only withholds the answer
维护者通常 1 天内回复
还没有人认领这个 Issue。
评估
调研方向
Start with the existing context checks in internal/frontend/grpc/runs.go and the handler locations listed in the issue, then trace the long parse, convert, query, render, and document operations. Add bounded cancellation checks through those paths and update the contract comment in client/opensysml/inprocess.go. Add deterministic counting-context tests for mid-loop cancellation and deadlines; done means each listed RPC returns the context status promptly and concurrent calls remain unaffected.
由索引模型根据 Issue 内容生成。
描述
Disclaimer, this issue was detected and written by an AI
Summary
Most engine calls honour a context: RunRuns, RunSweep, CheckBehavior,
CompareResults, Explore*, ImportData and SelfCheck notice ctx.Err() and
stop (or return it). The long parse/convert/query/render paths do not.
client/opensysml/inprocess.go:258 states the current contract:
The engine runs a call to completion; a done context withholds whatever it
produced, answer or error, as the wire does.
So a caller deadline returns DEADLINE_EXCEEDED while the engine keeps parsing,
converting or querying. On a memory-constrained host a handful of timed-out
requests stack full CPU and memory cost: the timeouts make the overload worse,
not better.
Where the checks already are
internal/frontend/grpc/runs.go:173,261(RunRuns)internal/frontend/grpc/sweep.go:143(RunSweep)internal/frontend/grpc/check.go:186(CheckBehavior)internal/frontend/grpc/compare.go:196(CompareResults)internal/frontend/grpc/explore.go:45(Explore*)internal/frontend/grpc/importdata.go:49(ImportData)internal/frontend/grpc/selfcheck.go:56(SelfCheck)
These return ctx.Err(); connect maps context.Canceled/DeadlineExceeded, and
the in-process answer maps them through contextStatus (inprocess.go:271).
Where they are missing (the work)
Add the same pattern — check at phase boundaries and every N elements/documents,
then return ctx.Err() — to:
| RPC | File | Where to check |
|---|---|---|
ParseFile |
internal/frontend/grpc/service.go:824 |
before/after the parse |
ParseSources |
service.go:857 |
per document in the set, and between analysis and the response build |
Convert |
export.go:28 |
before preparation, between export.PrepareModel, toGraph, graph build and serialization; per document in convertModelOfDocuments (export.go:564) |
Migrate |
export.go:189 |
per migrated document/phase |
Query |
query.go:99 |
in candidates/collect and when building rows, every ~1024 elements |
RenderView |
render_view.go |
before and after the render |
RenderDocument |
documents.go (render path) |
before and after the render |
ExportGraphs |
export_graphs.go |
before and after the export |
RunDocumentQuery |
docquery.go |
per row |
InstancePaths |
instancepaths.go |
per node during the walk |
ElementProperties |
element_properties.go |
per document/element |
Instantiate |
objects.go |
per behavior/object |
Notes:
- Check cheaply:
if err := ctx.Err(); err != nil { return nil, err }at the top
of each document/row and every few hundred elements; do not callctx.Err()
per property. - Return the context error unchanged, so the wire code is
CANCELLED/
DEADLINE_EXCEEDEDon both transports. Do not turn it into anINVALID_ARGUMENT
or a 200 witherror. - After this, update the
inprocess.go:258comment: the engine checks the
context where the work is long, and a done context returns the status. ParseSourcesis the biggest win in-process: a 572 KB/13-document project
takes ~2 s cold, and an abandoned tree read or export currently keeps it
running.
Tests
Service tests (internal/frontend/grpc), calling the handlers directly like
affected_test.go does:
- A counting context that answers nil for N
Err()calls and
context.Canceledafterwards.ParseSourcesandConvertover a large model
returnCodeCanceled(not a result), and return before the uncancelled call
would have finished. Pin the exact behaviour: with N below the first check the
call is untouched; with N inside the document loop it stops after the current
document. - The same for
Queryover a large scope. context.DeadlineExceededyieldsCodeDeadlineExceeded, notCodeCanceled.- A concurrent call on another model still completes (no shared cancellation).
- An in-process client test (
client/opensysml) drivesParseDocumentswith the
counting context and assertsStatusError{Code: CodeCanceled}and that the
partial model is not returned.
Avoid wall-clock-only tests: under a loaded 4-core box a 2 s parse is not a stable
oracle. The counting context makes the stop point deterministic.
Acceptance
- Every long RPC above stops within a bounded number of elements/documents of
a cancellation and returnsCANCELLED/DEADLINE_EXCEEDED. - A pre-cancelled request still returns immediately (unchanged).
- Concurrent calls are unaffected.
- The
answercomment and any docs that say "runs to completion" are updated. - Tests cover both cancellation and deadline, at a mid-loop check point.
- 主要语言
- Go
- 星标
- 59
- 派生
- 20
- 平均合并
- 10 小时 23 分钟
- 30 天内合并 PR
- 897
环境准备
- 提供 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
Open-MBEE/OpenSysML 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 76/100
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 35/100
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 35/100
维护者通常 1 天内回复
-
难度 4/5 3-5 天 新手友好度 55/100
维护者通常 1 天内回复
-
难度 5/5 一周以上 新手友好度 28/100
维护者通常 1 天内回复
查看 Open-MBEE/OpenSysML 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 78/100
维护者通常 1 天内回复
-
bug
难度 2/5 1-3 小时 新手友好度 66/100
AvengeMedia/DankMaterialShell#4301 · 1 条评论 ·
-
priority: p2 type: bug type: docs
难度 2/5 1-3 小时 新手友好度 84/100
googleapis/google-cloud-swift#1641 ·
维护者通常 1 天内回复
-
needs-triage
难度 2/5 1-3 小时 新手友好度 72/100
krakenhavoc/cmd_and_ctrl#5853 ·
维护者通常 1 天内回复
-
flaky test: TestTRTAsyncPrefillFailures/transport reads the token load before prefillDone has released it可能已有人在做 @Jackmeson1 今天认领。 未关闭area/cicd area/gateway kind/misc
难度 2/5 1-3 小时 新手友好度 85/100
vllm-project/aibrix#3488 · 2 条评论 ·
维护者通常 1 天内回复