fix(db): 路由跨 loop 的 DB 访问回引擎主 loop,消除 asyncmy "attached to a different loop"

会商/定时任务/沙箱工具等在 worker 线程的临时 loop 上用 asyncio.run 跑全局引擎连接池的
DB 操作,连接被绑定到随即关闭的临时 loop;池子回收/释放该连接时 asyncmy
_terminate_graceful_close 在错误的 loop 上触发 future,持续刷
"got Future attached to a different loop"。

- engine.py: 新增 run_db_blocking(fire_and_forget_db_write 的带返回值版本)。离开主 loop
  时用 run_coroutine_threadsafe 把协程送回主 loop 执行、阻塞调用线程拿结果;仅在无主 loop
  或正同步处于主 loop 内(送回会死锁)时退回临时 loop。
- thread_paths / prompt / skill_tools: 三处同步→异步阻塞读桥(任务owner解析/常驻技能/
  技能可见性)改走 run_db_blocking,删除各自重复的 _run_coro_blocking。
- llm/tool metrics 中间件: fire-and-forget 写改走 fire_and_forget_db_write。

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
吉鑫升andClaude Opus 4.8 committed 2026-06-24 15:57:04 +08:00
1 parent 5ecf28e60c
commit 6aff165d09
6 files changed
+147 -96

No files matched your search

@@ -86,27 +86,6 @@ def _drop_archived_skills(skills: list[Skill]) -> list[Skill]:
return skills
def _run_coro_blocking(coro):
"""Run an async coroutine to completion from sync code.
Mirrors ``skill_tools._run_coro_blocking``: when already inside a running
event loop (e.g. prompt built during an async agent run), hand the coroutine
to a fresh loop in a worker thread; otherwise run it directly.
"""
try:
asyncio.get_running_loop()
in_loop = True
except RuntimeError:
in_loop = False
if in_loop:
import concurrent.futures
with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
return pool.submit(asyncio.run, coro).result()
return asyncio.run(coro)
def _load_always_on_names() -> set[str]:
"""Names marked 常驻(免压缩), read from the **DB** (skills table).
@@ -114,11 +93,11 @@ def _load_always_on_names() -> set[str]:
simply treats every skill as compressible (safe default).
"""
try:
from deerflow.persistence.engine import get_session_factory
from deerflow.persistence.engine import get_session_factory, run_db_blocking
from deerflow.persistence.skills import make_skill_store
store = make_skill_store(get_session_factory())
return set(_run_coro_blocking(store.list_always_on_names()))
return set(run_db_blocking(store.list_always_on_names()))
except Exception:
logger.debug("Failed to load always_on skill names from DB", exc_info=True)
return set()
@@ -20,7 +20,6 @@ not the retry-wrapped successful response.
from __future__ import annotations
import asyncio
import logging
import time
import uuid
@@ -38,7 +37,7 @@ from langchain.agents.middleware.types import (
from langchain_core.messages import AIMessage
from langgraph.errors import GraphBubbleUp
from deerflow.persistence.engine import get_session_factory
from deerflow.persistence.engine import fire_and_forget_db_write, get_session_factory
from deerflow.persistence.llm_metrics import LlmMetricRecord, make_llm_metrics_store
from deerflow.persistence.llm_metrics.base import LlmMetricsStore
from deerflow.runtime.user_context import get_effective_user_id
@@ -280,16 +279,13 @@ class LlmMetricsMiddleware(AgentMiddleware[AgentState]):
store = self._store()
if store is None:
return
# Route the best-effort write onto the engine's home loop. Doing
# ``asyncio.run(store.record(...))`` here (when the model call ran in a worker
# thread) would reuse the global engine's pooled connections on a fresh loop and
# raise "got Future attached to a different loop" — harmless (caught below) but
# log-spammy and pool-churning under concurrency. See ``fire_and_forget_db_write``.
try:
try:
loop = asyncio.get_running_loop()
except RuntimeError:
loop = None
if loop is not None and loop.is_running():
# Fire-and-forget — never let metric IO block the request path.
loop.create_task(store.record(record))
else:
asyncio.run(store.record(record))
fire_and_forget_db_write(store.record(record))
except Exception:
logger.warning("LlmMetricsMiddleware: failed to flush metric %s", record.id, exc_info=True)
@@ -25,7 +25,6 @@ agent runtime keeps running.
from __future__ import annotations
import asyncio
import logging
import re
import time
@@ -41,7 +40,7 @@ from langgraph.errors import GraphBubbleUp
from langgraph.prebuilt.tool_node import ToolCallRequest
from langgraph.types import Command
from deerflow.persistence.engine import get_session_factory
from deerflow.persistence.engine import fire_and_forget_db_write, get_session_factory
from deerflow.persistence.tool_metrics import ToolMetricRecord, make_tool_metrics_store
from deerflow.persistence.tool_metrics.base import ToolMetricsStore
from deerflow.runtime.user_context import get_effective_user_id
@@ -262,16 +261,12 @@ class ToolMetricsMiddleware(AgentMiddleware[AgentState]):
store = self._store()
if store is None:
return
# Route the best-effort write onto the engine's home loop (a tool call that ran
# in a worker thread would otherwise do ``asyncio.run(store.record(...))`` and reuse
# the global engine's pooled connections cross-loop → "got Future attached to a
# different loop"). See ``fire_and_forget_db_write``.
try:
try:
loop = asyncio.get_running_loop()
except RuntimeError:
loop = None
if loop is not None and loop.is_running():
# Fire-and-forget — never let metric IO block the tool path.
loop.create_task(store.record(record))
else:
asyncio.run(store.record(record))
fire_and_forget_db_write(store.record(record))
except Exception:
logger.warning("ToolMetricsMiddleware: failed to flush metric %s", record.id, exc_info=True)