perf: 翻译独立后台循环(1 篇/秒)+ Semaphore 1

之前 fetch_one_source 入库后立即调翻译(可能并发触发腾讯 TMT 限速)
改为独立 translation_loop 后台循环:
- 完全不和 RSS 抓取并行
- 1 篇/秒节拍(Semaphore 1 + sleep 1.0)
- 没活时空闲 5 秒再轮询
- pending/failed 都重试

backend/app/workers/__main__.py

@@ -17,7 +17,7 @@ from sqlalchemy import select
 from app.config import settings
 from app.database import AsyncSessionLocal
 from app.models.source import Source
-from app.workers.pipeline import fetch_one_source, run_once
+from app.workers.pipeline import fetch_one_source, run_once, translation_loop
 
 logger = logging.getLogger("news.worker")
 logging.basicConfig(
@@ -89,6 +89,10 @@ async def main() -> None:
     scheduler.start()
     logger.info("scheduler started with %d jobs", len(scheduler.get_jobs()))
 
+    # 独立的翻译后台循环(不和 RSS 抓取并行;1 篇/秒)
+    translation_task = asyncio.create_task(translation_loop(), name="translation_loop")
+    logger.info("translation_loop task scheduled (1 article/sec)")
+
     stop = asyncio.Event()
 
     def _signal_handler():
@@ -104,7 +108,12 @@ async def main() -> None:
             pass
 
     await stop.wait()
-    logger.info("stopping scheduler")
+    logger.info("stopping scheduler and translation loop")
+    translation_task.cancel()
+    try:
+        await translation_task
+    except asyncio.CancelledError:
+        pass
     scheduler.shutdown(wait=False)
 
 

backend/app/workers/pipeline.py

@@ -51,10 +51,7 @@ async def fetch_one_source(source_id: int) -> None:
 
         n_new = await _bulk_insert(src, items)
         await _mark_success(source_id, n_new=n_new)
-        logger.info("source %s: %d new articles", src.slug, n_new)
-
-        # 入库后,挑高优先级 / 没翻译的开始翻译
-        await _translate_recent_for_source(source_id, max_n=20)
+        logger.info("source %s: %d new articles (translation deferred to background loop)", src.slug, n_new)
 
 
 async def _mark_failure(source_id: int, status: str) -> None:
@@ -271,3 +268,47 @@ async def run_once() -> None:
     logger.info("run_once: %d enabled sources", len(sources))
     tasks = [fetch_one_source(s.id) for s in sources]
     await asyncio.gather(*tasks, return_exceptions=True)
+
+
+# === 翻译后台循环 ===
+# 1 篇/秒(Semaphore 1 已经在 service 内部,这里是节拍)
+TRANSLATION_INTERVAL_SEC = 1.0
+TRANSLATION_IDLE_INTERVAL_SEC = 5.0
+TRANSLATION_BATCH_SIZE = 1  # 每轮最多翻译 1 篇
+
+
+async def translation_loop() -> None:
+    """独立的翻译 worker。
+    - 不和 RSS 抓取并行
+    - 1 篇/秒(用 TRANSLATION_INTERVAL_SEC 控制)
+    - 失败 status 写 'failed',下一次循环重试
+    """
+    logger.info("translation_loop started (interval=%.1fs)", TRANSLATION_INTERVAL_SEC)
+    while True:
+        try:
+            async with AsyncSessionLocal() as session:
+                rows = (
+                    await session.execute(
+                        select(Article)
+                        .where(Article.translation_status.in_(("pending", "failed")))
+                        .order_by(Article.fetched_at.asc(), Article.id.asc())
+                        .limit(TRANSLATION_BATCH_SIZE)
+                    )
+                ).scalars()
+                aids = [a.id for a in rows]
+
+            if not aids:
+                # 没活,等久一点
+                await asyncio.sleep(TRANSLATION_IDLE_INTERVAL_SEC)
+                continue
+
+            for aid in aids:
+                try:
+                    await translate_article(aid)
+                except Exception as e:
+                    logger.exception("translate_article %s failed: %s", aid, e)
+                # 1 篇/秒节拍
+                await asyncio.sleep(TRANSLATION_INTERVAL_SEC)
+        except Exception as e:
+            logger.exception("translation_loop error: %s", e)
+            await asyncio.sleep(TRANSLATION_IDLE_INTERVAL_SEC)