fix: 收敛 Agent 会话队列与 worker 关闭 (#6386)

This commit is contained in:
InfinityPacer
2026-08-22 10:09:01 +08:00
committed by GitHub
parent 813ea77bc4
commit 1c07333909
6 changed files with 656 additions and 31 deletions
+286 -27
View File
@@ -2448,6 +2448,7 @@ class _MessageTask:
agent_factory: Optional[Callable[..., MoviePilotAgent]] = None
agent_setup: Optional[Callable[[MoviePilotAgent], None]] = None
completion_future: Optional[asyncio.Future] = None
enqueued_at: Optional[float] = None
class AgentManagerUnavailableError(RuntimeError):
@@ -2456,6 +2457,23 @@ class AgentManagerUnavailableError(RuntimeError):
code = "agent_manager_unavailable"
class AgentManagerQueueFullError(RuntimeError):
"""Agent 会话的待处理消息达到容量上限。"""
code = "agent_manager_queue_full"
def __init__(self, session_id: str, limit: int):
self.session_id = session_id
self.limit = limit
super().__init__(
f"Agent 会话当前排队消息已达上限({limit} 条),请稍后重试"
)
AGENT_SESSION_QUEUE_MAX_SIZE = 8
AGENT_MANAGER_SHUTDOWN_TIMEOUT = 10.0
class AgentManager:
"""
AI智能体管理器
@@ -2473,6 +2491,14 @@ class AgentManager:
self._idle_cleanup_task: Optional[asyncio.Task] = None
self._idle_session_ttl = timedelta(hours=24)
self._idle_cleanup_interval = 60 * 60
self._session_queue_rejections: Dict[str, int] = {}
self._session_last_queue_wait_ms: Dict[str, float] = {}
self._session_shutdown_pending: Dict[str, asyncio.Task] = {}
self._session_cleanup_pending: set[str] = set()
self._session_deferred_cleanup_tasks: Dict[str, asyncio.Task] = {}
self._session_cancel_requested: set[str] = set()
self._close_finalizer_task: Optional[asyncio.Task] = None
self._shutdown_timeout = AGENT_MANAGER_SHUTDOWN_TIMEOUT
# 接收门禁与队列写入共用一把锁,确保关闭开始后不会再创建 worker。
self._lifecycle_lock = asyncio.Lock()
self._accepting_tasks = False
@@ -2494,6 +2520,20 @@ class AgentManager:
queue = self._session_queues.get(session_id)
status["pending_messages"] = queue.qsize() if queue else 0
status["queue_capacity"] = AGENT_SESSION_QUEUE_MAX_SIZE
status["queue_saturated"] = bool(queue and queue.full())
status["queue_rejections"] = self._session_queue_rejections.get(
session_id,
0,
)
status["last_queue_wait_ms"] = self._session_last_queue_wait_ms.get(
session_id,
0.0,
)
pending_shutdown = self._session_shutdown_pending.get(session_id)
status["shutdown_pending"] = bool(
pending_shutdown and not pending_shutdown.done()
)
status["is_processing"] = (
session_id in self._session_workers
and not self._session_workers[session_id].done()
@@ -2537,6 +2577,8 @@ class AgentManager:
关闭管理器
"""
async with self._lifecycle_lock:
if self._close_finalizer_task and not self._close_finalizer_task.done():
return
# 门禁必须先关闭;锁内完成清理可阻止等待中的请求在收口期间重新入队。
self._accepting_tasks = False
if self._idle_cleanup_task:
@@ -2546,16 +2588,20 @@ class AgentManager:
except asyncio.CancelledError:
pass
self._idle_cleanup_task = None
# 取消所有会话worker
for task in list(self._session_workers.values()):
# 取消所有 worker,再以有限等待收口,避免关闭阶段无限挂起。
workers = list(self._session_workers.items())
for session_id, task in workers:
self._session_cancel_requested.add(session_id)
task.cancel()
# 等待所有worker结束
for session_id, task in list(self._session_workers.items()):
try:
await task
except asyncio.CancelledError:
pass
self._session_workers.clear()
timed_out_workers = []
for session_id, task in workers:
stopped = await self._wait_for_worker_shutdown(
session_id,
task,
reason="manager_close",
)
if not stopped:
timed_out_workers.append((session_id, task))
for queue in list(self._session_queues.values()):
self._discard_queued_messages(
queue,
@@ -2563,6 +2609,32 @@ class AgentManager:
)
self._session_queues.clear()
self._session_last_used.clear()
self._session_queue_rejections.clear()
self._session_last_queue_wait_ms.clear()
if timed_out_workers:
timed_out_session_ids = {
session_id for session_id, _ in timed_out_workers
}
for session_id, task in workers:
if session_id in timed_out_session_ids:
continue
if self._session_workers.get(session_id) is task:
self._session_workers.pop(session_id, None)
for session_id, agent in list(self.active_agents.items()):
if session_id not in timed_out_session_ids:
await agent.cleanup()
self.active_agents.pop(session_id, None)
logger.error(
"AgentManager 关闭时仍有 worker 未收敛,"
f"保留 {len(timed_out_workers)} 个会话资源直到 worker 结束"
)
self._close_finalizer_task = asyncio.create_task(
self._finish_deferred_close(timed_out_workers)
)
return
self._session_workers.clear()
for agent in list(self.active_agents.values()):
await agent.cleanup()
self.active_agents.clear()
@@ -2670,16 +2742,45 @@ class AgentManager:
)
async with self._lifecycle_lock:
if not self._accepting_tasks:
if completion_future and not completion_future.done():
completion_future.cancel()
raise AgentManagerUnavailableError("AgentManager 未运行或已关闭")
pending_shutdown = self._session_shutdown_pending.get(session_id)
if pending_shutdown:
if pending_shutdown.done():
self._session_shutdown_pending.pop(session_id, None)
else:
if completion_future and not completion_future.done():
completion_future.cancel()
raise AgentManagerUnavailableError(
f"Agent 会话 {session_id} 仍在停止,暂时不能接收新任务"
)
self._record_session_activity(session_id, user_id)
# 获取或创建会话队列
if session_id not in self._session_queues:
self._session_queues[session_id] = asyncio.Queue()
self._session_queues[session_id] = asyncio.Queue(
maxsize=AGENT_SESSION_QUEUE_MAX_SIZE
)
queue = self._session_queues[session_id]
queue_size = queue.qsize()
if queue.full():
self._session_queue_rejections[session_id] = (
self._session_queue_rejections.get(session_id, 0) + 1
)
logger.warning(
f"会话 {session_id} 的 Agent 排队已满,拒绝新消息 "
f"(上限: {AGENT_SESSION_QUEUE_MAX_SIZE})"
)
if completion_future and not completion_future.done():
completion_future.cancel()
raise AgentManagerQueueFullError(
session_id=session_id,
limit=AGENT_SESSION_QUEUE_MAX_SIZE,
)
# 如果队列中已有等待的消息,通知用户消息已排队
if queue_size > 0 or (
session_id in self._session_workers
@@ -2690,8 +2791,9 @@ class AgentManager:
f"(队列中待处理: {queue_size} 条)"
)
# 放入队列并创建 worker 与关闭门禁保持原子关系
await queue.put(task)
# 非阻塞入队与 worker 创建在同一生命周期锁内完成,关闭期间不会留下悬挂入队
task.enqueued_at = asyncio.get_running_loop().time()
queue.put_nowait(task)
if (
session_id not in self._session_workers
or self._session_workers[session_id].done()
@@ -2724,10 +2826,29 @@ class AgentManager:
break
try:
if task.enqueued_at is not None:
queue_wait_ms = max(
0.0,
(
asyncio.get_running_loop().time()
- task.enqueued_at
)
* 1000,
)
self._session_last_queue_wait_ms[session_id] = round(
queue_wait_ms,
3,
)
await self._start_task_processing_status(task)
result = await self._process_message_internal(task)
if task.completion_future and not task.completion_future.done():
task.completion_future.set_result(result)
if (
not self._accepting_tasks
or session_id in self._session_cancel_requested
):
task.completion_future.cancel()
else:
task.completion_future.set_result(result)
except asyncio.CancelledError:
if task.completion_future and not task.completion_future.done():
if self._accepting_tasks:
@@ -2744,6 +2865,8 @@ class AgentManager:
finally:
await self._finish_task_processing_status(task)
queue.task_done()
if session_id in self._session_cancel_requested:
break
except asyncio.CancelledError:
logger.info(f"会话 {session_id} 的worker被取消")
@@ -2752,6 +2875,7 @@ class AgentManager:
current_worker = asyncio.current_task()
if self._session_workers.get(session_id) is current_worker:
self._session_workers.pop(session_id, None) # noqa
self._session_cancel_requested.discard(session_id)
# 如果队列为空,清理队列
if (
self._session_queues.get(session_id) is queue
@@ -2884,15 +3008,17 @@ class AgentManager:
# 先摘下旧队列再等待 worker 退出;lifecycle 锁保证清理期间不会并发建立新队列。
if worker:
self._session_cancel_requested.add(session_id)
worker.cancel()
if queue:
self._discard_queued_messages(queue)
if worker:
try:
await worker
except asyncio.CancelledError:
pass
if self._session_workers.get(session_id) is worker:
stopped_cleanly = await self._wait_for_worker_shutdown(
session_id,
worker,
reason="stop_current_task",
)
if stopped_cleanly and self._session_workers.get(session_id) is worker:
self._session_workers.pop(session_id, None) # noqa
stopped = True
if queue:
@@ -2905,9 +3031,10 @@ class AgentManager:
and not new_queue.empty()
and (not current_worker or current_worker.done())
):
self._session_workers[session_id] = asyncio.create_task(
self._session_worker(session_id)
)
if session_id not in self._session_shutdown_pending:
self._session_workers[session_id] = asyncio.create_task(
self._session_worker(session_id)
)
if stopped:
logger.info(f"会话 {session_id} 的Agent推理已应急停止")
@@ -2925,20 +3052,42 @@ class AgentManager:
async def _clear_session_locked(self, session_id: str, user_id: str) -> None:
"""在 lifecycle 互斥域内释放会话、Agent 与记忆。"""
if session_id in self._session_cleanup_pending:
return
self._session_last_used.pop(session_id, None)
# 取消该会话的worker
if session_id in self._session_workers:
self._session_workers[session_id].cancel()
try:
await self._session_workers[session_id]
except asyncio.CancelledError:
pass
self._session_workers.pop(session_id, None) # noqa
worker = self._session_workers[session_id]
self._session_cleanup_pending.add(session_id)
self._session_cancel_requested.add(session_id)
worker.cancel()
stopped_cleanly = await self._wait_for_worker_shutdown(
session_id,
worker,
reason="clear_session",
)
if not stopped_cleanly:
queue = self._session_queues.pop(session_id, None)
if queue:
self._discard_queued_messages(queue)
worker.add_done_callback(
lambda done: self._schedule_deferred_session_cleanup(
session_id,
user_id,
worker,
)
)
return
if self._session_workers.get(session_id) is worker:
self._session_workers.pop(session_id, None) # noqa
self._session_cleanup_pending.discard(session_id)
# 清理队列时同步结束未执行请求,避免 wait_for_completion 调用方永久等待。
queue = self._session_queues.pop(session_id, None)
if queue:
self._discard_queued_messages(queue)
self._session_queue_rejections.pop(session_id, None)
self._session_last_queue_wait_ms.pop(session_id, None)
# 清理agent
if session_id in self.active_agents:
@@ -2948,6 +3097,116 @@ class AgentManager:
memory_manager.clear_memory(session_id, user_id)
logger.info(f"会话 {session_id} 的记忆已清空")
def _schedule_deferred_session_cleanup(
self,
session_id: str,
user_id: str,
worker: asyncio.Task,
) -> None:
"""worker 超时后延迟释放会话资源,避免与仍在运行的 Agent 竞态。"""
if session_id in self._session_shutdown_pending:
cleanup_task = asyncio.create_task(
self._finish_deferred_session_cleanup(
session_id=session_id,
user_id=user_id,
worker=worker,
)
)
self._session_deferred_cleanup_tasks[session_id] = cleanup_task
async def _finish_deferred_session_cleanup(
self,
session_id: str,
user_id: str,
worker: asyncio.Task,
) -> None:
"""等待超时 worker 真正结束后,再完成 clear_session 的资源释放。"""
try:
await worker
except BaseException:
pass
async with self._lifecycle_lock:
if self._session_workers.get(session_id) is worker:
self._session_workers.pop(session_id, None)
self._session_shutdown_pending.pop(session_id, None)
self._session_cleanup_pending.discard(session_id)
self._session_deferred_cleanup_tasks.pop(session_id, None)
self._session_queue_rejections.pop(session_id, None)
self._session_last_queue_wait_ms.pop(session_id, None)
agent = self.active_agents.pop(session_id, None)
if agent:
await agent.cleanup()
memory_manager.clear_memory(session_id, user_id)
logger.info(f"会话 {session_id} 的记忆已清空")
async def _finish_deferred_close(
self,
workers: list[tuple[str, asyncio.Task]],
) -> None:
"""关闭超时后等待遗留 worker,再释放共享 Agent 资源。"""
try:
await asyncio.gather(
*(worker for _, worker in workers),
return_exceptions=True,
)
async with self._lifecycle_lock:
for session_id, worker in workers:
if self._session_workers.get(session_id) is worker:
self._session_workers.pop(session_id, None)
agent = self.active_agents.pop(session_id, None)
if agent:
await agent.cleanup()
for session_id, agent in list(self.active_agents.items()):
await agent.cleanup()
self.active_agents.pop(session_id, None)
self._session_shutdown_pending.clear()
self._session_cancel_requested.clear()
await memory_manager.close()
finally:
self._close_finalizer_task = None
async def _wait_for_worker_shutdown(
self,
session_id: str,
worker: asyncio.Task,
*,
reason: str,
) -> bool:
"""有限等待 worker 结束,超时会话保持停止态直到旧 worker 收敛。"""
try:
await asyncio.wait_for(
asyncio.shield(worker),
timeout=self._shutdown_timeout,
)
return True
except asyncio.CancelledError:
raise
except asyncio.TimeoutError:
self._session_shutdown_pending[session_id] = worker
def _clear_pending(done: asyncio.Task) -> None:
if (
self._session_shutdown_pending.get(session_id) is done
and session_id not in self._session_cleanup_pending
):
self._session_shutdown_pending.pop(session_id, None)
logger.info(
f"会话 {session_id} 的 Agent worker 已在超时后收敛"
)
worker.add_done_callback(_clear_pending)
logger.error(
f"会话 {session_id} 的 Agent worker 关闭超时,"
f"已阻止新任务进入,reason={reason}, timeout={self._shutdown_timeout:g}s"
)
return False
except Exception as error:
logger.error(
f"等待会话 {session_id} 的 Agent worker 关闭失败,"
f"reason={reason}: {error}"
)
return True
async def run_background_prompt(
self,
message: str,