Add Redis-backed chat queue

ai_knowledge_base_v2/apps/backend/app/api/chat.py

@@ -21,7 +21,13 @@ from app.schemas.chat import (
     UpdateSessionTitleRequest,
 )
 from app.services.chat_service import ChatService
-from app.services.chat_queue_service import chat_queue_manager, load_chat_queue_config
+from app.services.chat_queue_runtime import (
+    cancel_chat_waiter,
+    release_chat_slot,
+    request_chat_slot,
+    wait_for_chat_slot,
+)
+from app.services.chat_queue_service import load_chat_queue_config
 from app.services.chat_stream_service import ChatStreamService
 
 router = APIRouter()
@@ -90,9 +96,10 @@ def stop(
 
 async def _chat_stream(payload: ChatCompletionRequest, db: Session, current_user: User) -> AsyncIterator[str]:
     config = load_chat_queue_config(db)
-    queue_request = chat_queue_manager.request_slot(config)
+    queue_lease = await request_chat_slot(config)
+    queue_request = queue_lease.request
     acquired = queue_request.status == "acquired"
-    queued_token = queue_request.token
+    should_cancel_waiter = queue_request.status == "queued"
 
     if queue_request.status == "rejected":
         yield _sse_event(
@@ -112,12 +119,27 @@ async def _chat_stream(payload: ChatCompletionRequest, db: Session, current_user
             activeCount=queue_request.active_count,
             waitingCount=queue_request.waiting_count,
         )
-        queue_request = await asyncio.to_thread(chat_queue_manager.wait_for_slot, queued_token, config)
+        queue_lease = await wait_for_chat_slot(queue_lease, config)
+        queue_request = queue_lease.request
         acquired = queue_request.status == "acquired"
+        should_cancel_waiter = queue_request.status == "queued"
+        if queue_request.status == "rejected":
+            yield _sse_event(
+                "error",
+                message="当前请求过多，请稍后再试。",
+                activeCount=queue_request.active_count,
+                waitingCount=queue_request.waiting_count,
+            )
+            yield _sse_done()
+            return
         if queue_request.status == "timeout":
             yield _sse_event("error", message="排队等待超时，请稍后再试。")
             yield _sse_done()
             return
+        if queue_request.status == "canceled":
+            yield _sse_event("error", message="排队请求已取消，请稍后再试。")
+            yield _sse_done()
+            return
 
     try:
         yield _sse_event(
@@ -136,9 +158,9 @@ async def _chat_stream(payload: ChatCompletionRequest, db: Session, current_user
         yield _sse_event("error", message="AI 回复生成失败，请稍后再试。")
     finally:
         if acquired:
-            chat_queue_manager.release_slot()
-        elif queued_token is not None:
-            chat_queue_manager.cancel_waiter(queued_token)
+            await release_chat_slot(queue_lease)
+        elif should_cancel_waiter:
+            await cancel_chat_waiter(queue_lease)
 
     yield _sse_done()