引擎四处:错误处理、恢复语义、幂运算炸弹

## 一条分支条件能把后端冻死

求值器放行 ast.Pow,唯一护栏是 AST 节点数 <= 200。`9**9**9**9` 只有 11 个节点,
却让 CPython 无限期算大整数。而求值是在事件循环里同步执行的(分支条件、
transform、口径卡、每个节点的 skip_if),没有线程池隔离也没有超时,asyncio 的
cancel 打断不了同步 CPU。给底数和指数加上界,超界报 ExpressionError。
2**10、2**0.5、10**-2 这些正常用法不受影响。

## NodeError 绕过 retries / on_error / node.failed

`except NodeError: raise` 排在重试与兜底之前,于是节点自己抛的错既不重试、
也不发 node.failed、更不被 on_error=continue 接住——而 NodeError 恰恰是所有
执行器表达失败的标准方式(缺必填输入、模型没配、口径卡算不出…)。同一张图
配着同样的 retries=1 + on_error=continue,1/0 会重试并继续,而 JSON 模板非法
直接整个 run 失败。两条路径现在行为一致,node.failed 也带上 node_id
(前端靠它把出错节点从"转圈"收敛成"失败")。

## 没有 checkpoint 的运行被"恢复"成一份假结果

重启扫描把 running 和 queued 一律标成 interrupted,而 resume 不检查有没有
checkpoint 就发 Command。queued 的运行从没写过 checkpoint,LangGraph 会拿空
state 从 START 重跑整张图——run.input 完全没注入,用字段默认值算完还落成
succeeded。现在 queued 直接标 failed(没有进度可恢复,请重新发起),resume
先 aget_state 确认有断点,没有就 409 并说清为什么。

## 并行人工节点:一次 resume 打爆整个 run

同一 superstep 有多个 interrupt 时,resume 仍只发标量 Command,LangGraph 直接
抛 RuntimeError;而所有 pending 审批已经被写成同一个回复。Approval.interrupt_id
和 ResumeIn.approval_id 这两个字段一直存在却从没被用过。现在按 interrupt_id
逐条收集:多条待办时必须指定 approval_id,回复的那条转 answered 继续等,
凑齐引擎在等的全部 interrupt 才发 Command(resume={iid: value})。

验证:两个并行 human 节点——不指定 approval_id 被 409 拒、回复第一条后仍
interrupted 且剩 1 条 pending、凑齐后 succeeded。单人审批与 decide 通道回归通过。
This commit is contained in:
ThearchyHelios
2026-09-14 15:35:47 +08:00
parent 0afffd4b9a
commit b0575a7d1b
4 changed files with 170 additions and 24 deletions
+15 -3
View File
@@ -229,9 +229,17 @@ class ResumeIn(BaseModel):
@router.post("/{run_id}/resume", response_model=RunOut)
async def resume_run(run_id: str, payload: ResumeIn) -> Run:
async def resume_run(
run_id: str, payload: ResumeIn, x_actor: str | None = Header(default=None)
) -> Run:
try:
return await run_manager.resume(run_id, payload.response)
# approval_id 以前定义了却从没传下去——一次运行里有多条待审批时,
# 引擎无从知道这次回复的是哪一条
return await run_manager.resume(
run_id, payload.response,
approval_id=payload.approval_id,
actor=(x_actor or "").strip() or None,
)
except KeyError as e:
raise HTTPException(404, str(e)) from e
except ValueError as e:
@@ -443,7 +451,11 @@ async def decide(
response["args"] = payload.args
try:
return await run_manager.resume(approval.run_id, response)
return await run_manager.resume(
approval.run_id, response,
approval_id=approval.id,
actor=approval.resolved_by,
)
except ValueError as e:
raise HTTPException(409, str(e)) from e
+21 -5
View File
@@ -114,23 +114,26 @@ def _wrap(node: GraphNode, run_ctx: RunContext) -> Callable[[GraphState], Awaita
raise
except asyncio.CancelledError:
raise
except NodeError:
raise
except Exception as e: # noqa: BLE001
# NodeError 以前在这里直接 raise,于是节点自己抛的错既不重试、
# 也不发 node.failed、更不被 on_error=continue 接住。而 NodeError
# 恰恰是所有执行器表达失败的标准方式(缺必填输入、模型没配、
# 口径卡算不出…),等于这两个配置对最常见的报错全都不生效。
last_error = e
if attempt < retries:
ctx.emit(
EventType.LOG,
level="warn",
message=f"第 {attempt + 1} 次失败({type(e).__name__}: {e}),准备重试",
message=f"第 {attempt + 1} 次失败({_describe(e)}),准备重试",
)
await asyncio.sleep(backoff * (2**attempt))
continue
break
elapsed = int((time.perf_counter() - started) * 1000)
message = f"{type(last_error).__name__}: {last_error}"
ctx.emit(EventType.NODE_FAILED, error=message, duration_ms=elapsed)
message = _describe(last_error)
# node_id 要带上:前端靠它把出错的节点从"转圈"收敛成"失败"
ctx.emit(EventType.NODE_FAILED, error=message, duration_ms=elapsed, node_id=node.id)
# 容错模式:记下错误继续往下走,而不是让整张图挂掉
if node.config.get("on_error") == "continue":
@@ -138,6 +141,10 @@ def _wrap(node: GraphNode, run_ctx: RunContext) -> Callable[[GraphState], Awaita
"nodes": {node.id: {"error": message, "failed": True}},
"trail": [{"node_id": node.id, "error": message, "ts": time.time()}],
}
# 已经是 NodeError 就别再套一层,否则消息会变成
# "NodeError: NodeError: 真正的原因"
if isinstance(last_error, NodeError):
raise last_error
raise NodeError(node.id, message) from last_error
async def _entry(state: GraphState) -> dict[str, Any]:
@@ -146,6 +153,15 @@ def _wrap(node: GraphNode, run_ctx: RunContext) -> Callable[[GraphState], Awaita
return _entry
def _describe(error: BaseException | None) -> str:
"""错误消息。NodeError 的 message 本来就是写给人看的,不用再前缀类型名。"""
if error is None:
return "未知错误"
if isinstance(error, NodeError):
return str(error)
return f"{type(error).__name__}: {error}"
def _preview(updates: dict[str, Any], node_id: str) -> Any:
"""给前端卡片用的输出摘要,避免把整坨数据塞进事件流。"""
payload = (updates.get("nodes") or {}).get(node_id)
+22 -1
View File
@@ -107,6 +107,27 @@ def render_deep(value: Any, ctx: dict[str, Any]) -> Any:
# 安全表达式求值(分支条件用)
# --------------------------------------------------------------------------
# 幂运算的上界。`9**9**9**9` 只有 11 个 AST 节点,轻松过掉节点数护栏,
# 却会让 CPython 无限期地算一个天文数字的大整数 —— 而求值是在事件循环里
# 同步执行的(分支条件、transform、口径卡、每个节点的 skip_if 都走这里),
# 既没有线程池隔离也没有超时,asyncio 的 cancel 打断不了同步 CPU。
# 一条画布上的分支条件就能把整个后端冻死。
_MAX_POW_EXPONENT = 1024
_MAX_POW_BASE = 1e15
def _safe_pow(base: Any, exponent: Any) -> Any:
"""带上界的幂运算:超界直接报错,而不是让进程算到天荒地老。"""
try:
if abs(exponent) > _MAX_POW_EXPONENT:
raise ExpressionError(f"指数超过上限 {_MAX_POW_EXPONENT}")
if abs(base) > _MAX_POW_BASE and exponent > 1:
raise ExpressionError(f"底数超过上限 {_MAX_POW_BASE:g}")
except TypeError as e: # 非数值类型交给 operator.pow 自己报
raise ExpressionError(f"幂运算的操作数必须是数值:{e}") from e
return operator.pow(base, exponent)
_BIN_OPS = {
ast.Add: operator.add,
ast.Sub: operator.sub,
@@ -114,7 +135,7 @@ _BIN_OPS = {
ast.Div: operator.truediv,
ast.FloorDiv: operator.floordiv,
ast.Mod: operator.mod,
ast.Pow: operator.pow,
ast.Pow: _safe_pow,
}
_CMP_OPS = {
ast.Eq: operator.eq,
+112 -15
View File
@@ -42,13 +42,21 @@ class RunManager:
settings.ensure_dirs()
self._cm = AsyncSqliteSaver.from_conn_string(str(settings.checkpoint_path))
self._checkpointer = await self._cm.__aenter__()
# 上次进程被强杀时留下的 running 记录,标成中断态而不是永远转圈
# 上次进程被强杀时留下的记录。running 和 queued 要分开处理:
# running 至少跑过一步、有 checkpoint,可以续;queued 还在排队,
# 一个 checkpoint 都没写过,"恢复"它只会从 START 用默认值重跑一遍,
# 还丢掉 run.input——那不是恢复,是伪造一份结果。
async with SessionLocal() as session:
await session.execute(
update(Run)
.where(Run.status.in_(["running", "queued"]))
.where(Run.status == "running")
.values(status="interrupted", error="服务重启,运行已挂起,可从断点恢复")
)
await session.execute(
update(Run)
.where(Run.status == "queued")
.values(status="failed", error="服务重启时这次运行还在排队,没有可恢复的进度,请重新发起")
)
await session.commit()
async def shutdown(self) -> None:
@@ -160,8 +168,23 @@ class RunManager:
)
return run
async def resume(self, run_id: str, response: Any) -> Run:
"""从人工介入的断点继续。"""
async def resume(
self, run_id: str, response: Any, *, approval_id: str | None = None,
actor: str | None = None,
) -> Run:
"""从人工介入的断点继续。
两件事以前是错的:
一是不看有没有 checkpoint 就发 Command。重启扫描会把 queued(从没跑过、
没有任何 checkpoint)也标成 interrupted,对这种 run 发 Command,LangGraph
会拿一个空 state 从 START 重跑整张图——run.input 完全没被注入,最后还落成
succeeded。跑出来的是一份用默认值算的假结果。
二是同一 superstep 里有多个 interrupt 时,只发标量 Command 会让 LangGraph
直接抛 RuntimeError 打爆整个 run,而所有 pending 审批还都被写成同一个回复。
现在按 interrupt_id 逐条收集,凑齐了才继续,没凑齐就继续等下一个人回复。
"""
async with SessionLocal() as session:
run = await session.get(Run, run_id)
if not run:
@@ -169,29 +192,103 @@ class RunManager:
if run.status not in ("interrupted",):
raise ValueError(f"当前状态是 {run.status},无法恢复")
spec = GraphSpec.model_validate(run.graph)
bus.set_seq(run_id, run.last_seq)
run.status = "running"
run.error = None
workflow_id = run.workflow_id
# 把这次恢复对应的待审批记录结掉,前端的待办列表才不会一直挂着
pending = (
# 先问引擎:这条线程到底停在哪、还等着哪些 interrupt
run_ctx = RunContext(run_id=run_id, thread_id=run_id, spec=spec, workflow_id=workflow_id)
app = compile_graph(spec, run_ctx).compile(checkpointer=self.checkpointer)
config = {"configurable": {"thread_id": run_id}}
snapshot = await app.aget_state(config)
waiting = list(getattr(snapshot, "interrupts", None) or [])
if not waiting and not getattr(snapshot, "next", None):
# 没有断点可续。多半是重启时被标成 interrupted 的 queued 运行。
raise ValueError(
"这次运行没有可恢复的断点(很可能重启前还没真正开始跑)。"
"请重新发起一次运行,而不是恢复——继续下去只会用默认值跑出一份假结果。"
)
answers: dict[str, Any] = {}
async with SessionLocal() as session:
pending = list((
await session.execute(
select(Approval).where(
Approval.run_id == run_id, Approval.status == "pending"
)
)
).scalars()
for approval in pending:
approval.status = "resolved"
approval.response = _safe(response) if isinstance(response, dict) else {"value": _safe(response)}
approval.resolved_at = datetime.now(timezone.utc)
).scalars())
# 定位这次回复的是哪一条:显式指定优先,否则只在唯一 pending 时才敢猜
target = None
if approval_id:
target = next((a for a in pending if a.id == approval_id), None)
if target is None:
raise ValueError(f"审批 {approval_id} 不在待处理列表里")
elif len(pending) == 1:
target = pending[0]
elif len(pending) > 1:
raise ValueError(
f"这次运行有 {len(pending)} 条待处理的人工介入,"
"恢复时必须指定 approval_id 说明回复的是哪一条"
)
now = datetime.now(timezone.utc)
if target is not None:
target.status = "answered" # 已回复,但还没提交给引擎
target.response = (
_safe(response) if isinstance(response, dict) else {"value": _safe(response)}
)
target.resolved_at = now
if actor:
target.resolved_by = actor
await session.commit()
# 凑齐没有?引擎还在等的每一个 interrupt 都得有答案
answered = list((
await session.execute(
select(Approval).where(
Approval.run_id == run_id,
Approval.status.in_(["answered", "pending"]),
)
)
).scalars())
by_iid = {a.interrupt_id: a for a in answered if a.interrupt_id}
missing = []
for item in waiting:
iid = getattr(item, "id", None)
rec = by_iid.get(iid)
if rec is None or rec.status != "answered":
missing.append(iid)
continue
payload = rec.response or {}
answers[iid] = payload.get("value", payload) if isinstance(payload, dict) else payload
if missing:
# 还差人没回。保持 interrupted,把已回复的那条留在 answered 上等齐
await session.commit()
run = await session.get(Run, run_id)
return run
# 齐了:正式落 resolved,然后驱动引擎
for rec in answered:
if rec.status == "answered":
rec.status = "resolved"
run = await session.get(Run, run_id)
bus.set_seq(run_id, run.last_seq)
run.status = "running"
run.error = None
await session.commit()
await session.refresh(run)
# 单个 interrupt 用标量(LangGraph 两种都认),多个必须用 {interrupt_id: value} 映射
command = (
Command(resume=response)
if len(waiting) <= 1
else Command(resume=answers)
)
await self._emit(run_id, EventType.RUN_RESUMED, data={"response": _safe(response)})
self._tasks[run_id] = asyncio.create_task(
self._drive(run_id, spec, Command(resume=response), workflow_id=run.workflow_id)
self._drive(run_id, spec, command, workflow_id=workflow_id)
)
return run