diff --git a/pipeline_service/agent_loop.py b/pipeline_service/agent_loop.py
index 1076e39..aef088d 100644
--- a/pipeline_service/agent_loop.py
+++ b/pipeline_service/agent_loop.py
@@ -941,6 +941,8 @@ async def handle_failed_task(task_id: str, project_id: str) -> dict:
             "WHERE id=${tid}$ AND state='failed'",
             {"tid": task_id})
         if not recs:
+            # 释放 SELECT 的元数据锁，防止连接回池后阻塞 DDL
+            await sor.sqlExe("COMMIT", {})
             return {"status": "idle", "task_id": task_id}
         t = recs[0]
         title = getattr(t, "title", "") or ""
@@ -951,6 +953,9 @@ async def handle_failed_task(task_id: str, project_id: str) -> dict:
             retry_count = 0
         last_error = getattr(t, "last_error", "") or ""
 
+        # 释放 SELECT 的元数据锁，避免决策/建问题期间长时间持有 MDL
+        await sor.sqlExe("COMMIT", {})
+
         # 已重试 3 次仍未成功 → 报故障
         if retry_count >= 3:
             decision = "fault"
diff --git a/pipeline_service/init.py b/pipeline_service/init.py
index 9f49193..bc450c3 100644
--- a/pipeline_service/init.py
+++ b/pipeline_service/init.py
@@ -663,6 +663,9 @@ def load_pipeline_service():
                                         _fd_dispatched.discard(tid)
 
                                 asyncio.ensure_future(_fd_dispatch(tid, pid))
+                        # SELECT-only 事务必须显式 COMMIT，否则连接回到池后仍持有元数据锁(MDL)，
+                        # 会阻塞后续 ALTER TABLE / DDL（sqlor 只在有写入时才自动提交）。
+                        await sor.sqlExe("COMMIT", {})
                 except Exception as e:
                     debug(f"failed_poller error: {e}")
                 await asyncio.sleep(60)