Transforms complex documents like PDFs and Office docs into LLM-ready markdown/JSON for your Agentic workflows.
Python ai4sciencedocument-analysisdocxextract-datalayout-analysisocrparserpdfpdf-converterpdf-extractor-llmpdf-extractor-pretrainpdf-extractor-ragpdf-parserpptxpythonxlsx
Reject stale worker responses and classify fatal preload failures; Refs #5188
myhloli6a5e7b2d9f9c
diff --git a/mineru/kit/router/app.py b/mineru/kit/router/app.pyindex 95f255d4..88a8c3c2 100644--- a/mineru/kit/router/app.py+++ b/mineru/kit/router/app.py@@ -268,6 +268,7 @@ async def _hydrate_job_output_files(if route is None or isinstance(route.metadata.get("payload"), dict):continueworker = pool.get(route.worker_id)+ generation = worker.generationtry:response = await request_upstream(pool,@@ -291,6 +292,8 @@ async def _hydrate_job_output_files("message": exc.message,}continue+ if worker.generation != generation or registry.find("file", public_file_id) is not route:+ continuerewrite_file_payload(metadata, worker, registry, route.owner_scope)route.metadata.pop("hydration_error", None)@@ -338,6 +341,7 @@ async def _reconcile_jobs_once(registry: ResourceRegistry, pool: WorkerPool) ->if not route.metadata.get("active_counted") and not route.metadata.get("copied_inputs"):continueworker = pool.get(route.worker_id)+ generation = worker.generationheaders = dict(route.metadata.get("upstream_headers") or {})try:response = await request_upstream(@@ -347,6 +351,8 @@ async def _reconcile_jobs_once(registry: ResourceRegistry, pool: WorkerPool) ->f"/v1/parse/jobs/{route.upstream_id}",headers=headers,)+ if worker.generation != generation or route.metadata.get("upstream_lost"):+ continueif response.status_code >= 400:cached = route.metadata.get("payload") or {}if response.status_code == 404 and cached.get("status") in {"completed", "partial", "failed", "canceled"}:@@ -963,7 +969,10 @@ def create_app(if route.metadata.get("upstream_lost"):return JSONResponse(route.metadata["payload"])worker = pool.get(route.worker_id)+ generation = worker.generationupstream = await request_upstream(pool, worker, "GET", f"/v1/parse/jobs/{route.upstream_id}", request=request)+ if worker.generation != generation or route.metadata.get("upstream_lost"):+ return JSONResponse(route.metadata["payload"])result = _successful_json(upstream)if isinstance(result, Response):return result@@ -975,6 +984,8 @@ def create_app(pool=pool,registry=registry,)+ if worker.generation != generation or route.metadata.get("upstream_lost"):+ return JSONResponse(route.metadata["payload"])return JSONResponse(rewritten, status_code=upstream.status_code)@application.delete("/v1/parse/jobs/{job_id}")diff --git a/mineru/kit/router/proxy.py b/mineru/kit/router/proxy.pyindex 385cf20e..3a701458 100644--- a/mineru/kit/router/proxy.py+++ b/mineru/kit/router/proxy.py@@ -80,6 +80,7 @@ async def request_upstream(follow_redirects: bool = False,) -> httpx.Response:"""执行普通 upstream 请求,并把连接失败与超时转成稳定 Router 错误。"""+ generation = worker.generationoutgoing_headers = forwarded_headers(request) if request is not None else {}outgoing_headers.pop("expect", None)if json_body is not None:@@ -94,7 +95,7 @@ async def request_upstream(headers=outgoing_headers,follow_redirects=follow_redirects,)- if response.status_code == 503:+ if response.status_code == 503 and worker.generation == generation:try:if (response.json().get("error") or {}).get("code") == "engine_dead":pool.notify_engine_dead(worker)diff --git a/mineru/parser/api_server.py b/mineru/parser/api_server.pyindex c09558e3..32239c1b 100644--- a/mineru/parser/api_server.py+++ b/mineru/parser/api_server.py@@ -2469,6 +2469,8 @@ _MODEL_PRELOAD_DEVICE_ERROR_PATTERNS = (def _classify_model_preload_error(exc: Exception) -> tuple[str, str]:message = str(exc).strip() or type(exc).__name__+ if is_fatal_engine_error(exc):+ return "engine_dead", messageif isinstance(exc, (ModuleNotFoundError, ImportError)):return "model_preload_dependency_missing", messageif isinstance(exc, FileNotFoundError):@@ -2654,6 +2656,8 @@ def create_app()except Exception as exc:error_code, error_msg = _classify_model_preload_error(exc)+ if error_code == "engine_dead":+ job_store.fail_engine(error_msg)application.state.model_preload_error = _ModelPreloadError(code=error_code,message=error_msg,diff --git a/tests/unittest/test_engine_dead_lifecycle.py b/tests/unittest/test_engine_dead_lifecycle.pyindex a2e99a4c..2c5f2437 100644--- a/tests/unittest/test_engine_dead_lifecycle.py+++ b/tests/unittest/test_engine_dead_lifecycle.py@@ -180,3 +180,41 @@ def test_old_generation_failure_snapshot_remains_queryable() -> None:assert response.json()["status"] == "failed"assert response.json()["files"][0]["error"]["code"] == "engine_dead"assert client.delete(f"/v1/parse/jobs/{route.public_id}").status_code == 409+++def test_old_generation_engine_dead_response_does_not_kill_replacement() -> None:+ """旧请求迟到报告死亡时,不得把健康的新 generation 再次标记死亡。"""+ from mineru.kit.router.proxy import request_upstream++ async def scenario() -> None:+ """响应处理之前模拟替换已发布,验证 notify 的 generation 条件。"""+ state = workers.WorkerState(+ "local",+ "http://old",+ "local",+ generation=1,+ healthy=True,+ local_worker=SimpleNamespace(process=SimpleNamespace(poll=lambda: None)),+ )++ def handle(request: httpx.Request) -> httpx.Response:+ """旧网络请求返回前发布新代,随后返回旧代的 engine_dead。"""+ state.generation = 2+ state.base_url = "http://new"+ state.healthy = True+ return httpx.Response(503, json={"error": {"code": "engine_dead"}})++ pool = workers.WorkerPool(workers.RouterSettings(local_gpus="none"), transport=httpx.MockTransport(handle))+ try:+ response = await request_upstream(pool, state, "GET", "/v1/health")+ assert response.status_code == 503+ assert state.healthy and not state.restart_requested+ finally:+ await pool.client.aclose()++ asyncio.run(scenario())+++def test_fatal_preload_uses_engine_dead_code() -> None:+ """预加载期间死亡与推理期间死亡共享同一稳定错误码。"""+ assert api_server._classify_model_preload_error(EngineDeadError("dead")) == ("engine_dead", "dead")