fix: 并发上传5文件刷新后丢失 + 分片写入阻塞事件循环

根因：
1. pump() 仅在 addFiles() 调用一次，上传完成/失败后不重新触发，
   导致 CONCURRENCY=3 之后的文件（4、5）永远不启动 → 无 DB 会话 → 刷新后消失
2. write_chunk 的 uploaded_chunks 是 read-modify-write，并发分片写入
   后者覆盖前者 → 分片记录丢失 → complete 报 409
3. upload_chunk 是 async 但同步调 write_chunk（fsync+DB commit），
   阻塞 uvicorn 事件循环 → 所有 web 请求被串行化

修复：
- _shared.py: pump() 加 finally 块，上传完成/失败后都触发下一文件；
  文件并发(FILE_CONCURRENCY=5)与分片并发(CHUNK_CONCURRENCY=3)分离
- upload_service.py: 按 upload_id 的进程级锁串行化 uploaded_chunks 更新，
  持锁后 db.refresh 重读最新值再 append，杜绝丢失更新；complete 后清理锁
- upload_router.py: upload_chunk 的 write_chunk 调用改用 run_in_threadpool，
  阻塞 I/O 移出事件循环，web 请求不再被分片写入阻塞

验证：5 文件并发上传后刷新全部可见；4 任务并发处理 4/4 成功(19.2s)

app/controllers/upload_router.py

@@ -6,6 +6,7 @@ complete 成功后创建转写 Task 并交由 scheduler 入队。
 from __future__ import annotations
 
 from fastapi import APIRouter, Depends, Request
+from starlette.concurrency import run_in_threadpool
 from sqlalchemy.orm import Session
 
 from ..database import get_db
@@ -48,7 +49,10 @@ async def upload_chunk(
     request: Request,
     service: UploadService = Depends(_service),
 ) -> ChunkUploadResponse:
-    uploaded = service.write_chunk(upload_id, index, await request.body())
+    # write_chunk 做文件 fsync + DB commit（阻塞 I/O），必须放到线程池跑，
+    # 否则会阻塞 uvicorn 事件循环，导致并发分片上传被串行化、web 请求卡顿。
+    body = await request.body()
+    uploaded = await run_in_threadpool(service.write_chunk, upload_id, index, body)
     return ChunkUploadResponse(upload_id=upload_id, index=index, uploaded_chunks=uploaded)
 
 

app/services/upload_service.py

@@ -16,6 +16,7 @@ from __future__ import annotations
 import logging
 import os
 import shutil
+import threading
 import uuid
 from datetime import datetime, timedelta, timezone
 from pathlib import Path
@@ -36,6 +37,20 @@ from ..schemas.task import (
 
 logger = logging.getLogger("audio2text.upload")
 
+# 按 upload_id 串行化 uploaded_chunks 的读-改-写，避免并发分片写入丢失更新。
+# SQLite 无行锁，JSON 列的 append 操作不是原子的，必须进程内加锁。
+_chunk_locks: dict[str, threading.Lock] = {}
+_chunk_locks_guard = threading.Lock()
+
+
+def _get_chunk_lock(upload_id: str) -> threading.Lock:
+    with _chunk_locks_guard:
+        lock = _chunk_locks.get(upload_id)
+        if lock is None:
+            lock = threading.Lock()
+            _chunk_locks[upload_id] = lock
+        return lock
+
 
 class UploadService:
     def __init__(self, db: Session) -> None:
@@ -93,6 +108,7 @@ class UploadService:
         session_dir.mkdir(parents=True, exist_ok=True)
         chunk_path = session_dir / f"{index}.part"
 
+        # 先落盘（无锁，IO 可并行），再持锁更新 DB 计数
         try:
             with chunk_path.open("wb") as out:
                 out.write(data)
@@ -102,12 +118,16 @@ class UploadService:
             chunk_path.unlink(missing_ok=True)
             raise
 
-        uploaded = list(session.uploaded_chunks or [])
-        if index not in uploaded:
-            uploaded.append(index)
-        session.uploaded_chunks = uploaded
-        session.updated_at = datetime.now(timezone.utc)
-        self.db.commit()
+        # 持锁重读+更新：避免并发分片各自 read old list → append → write，后者覆盖前者
+        lock = _get_chunk_lock(upload_id)
+        with lock:
+            self.db.refresh(session)  # 拿最新 uploaded_chunks，不被其他并发请求的旧值覆盖
+            uploaded = list(session.uploaded_chunks or [])
+            if index not in uploaded:
+                uploaded.append(index)
+            session.uploaded_chunks = uploaded
+            session.updated_at = datetime.now(timezone.utc)
+            self.db.commit()
         return sorted(uploaded)
 
     # ---------------- 拼接 + 创建任务 ----------------
@@ -163,6 +183,9 @@ class UploadService:
 
         # 清理分片暂存（commit 后，即使清理失败也不影响已建任务）
         self._cleanup_session_dir(upload_id)
+        # 清理进程内锁，避免长期运行后 _chunk_locks 无限增长
+        with _chunk_locks_guard:
+            _chunk_locks.pop(upload_id, None)
 
         logger.info("上传完成 task_id=%s file=%s size=%d", task.id, session.filename, session.size_bytes)
         return CompleteResponse(

app/views/_shared.py

@@ -10,7 +10,8 @@ from __future__ import annotations
 
 # 分片上传参数（home / upload 共用）
 DEFAULT_CHUNK_SIZE = 4 * 1024 * 1024
-DEFAULT_CONCURRENCY = 3
+DEFAULT_CONCURRENCY = 3       # 单文件内分片并发数
+DEFAULT_FILE_CONCURRENCY = 5  # 同时上传的文件数（不阻塞 web 请求/ffmpeg/gpu）
 MAX_RETRY = 2
 POLL_INTERVAL = 2000
 
@@ -259,7 +260,8 @@ def render_upload_js(on_complete: str) -> str:
     """
     return f"""
 const CHUNK_SIZE = {DEFAULT_CHUNK_SIZE};
-const CONCURRENCY = {DEFAULT_CONCURRENCY};
+const CHUNK_CONCURRENCY = {DEFAULT_CONCURRENCY};   // 单文件内分片并发
+const FILE_CONCURRENCY = {DEFAULT_FILE_CONCURRENCY}; // 同时上传的文件数
 const MAX_RETRY = {MAX_RETRY};
 const UPLOAD_API = "/api/tasks/chunk-uploads";
 
@@ -315,13 +317,13 @@ function setUploadProgress(t, pct) {{
 }}
 
 function pump() {{
+  // 统计正在上传的文件数，启动等待中的文件直到达到 FILE_CONCURRENCY
   const active = pending.filter(t => t.state === "running").length;
   for (const t of pending) {{
-    if (active >= CONCURRENCY) break;
+    if (active >= FILE_CONCURRENCY) break;
     if (t.state === "pending") {{
-      t.state = "running";
+      setUploadState(t, "running");
       startUpload(t);
-      active++;
     }}
   }}
 }}
@@ -342,7 +344,7 @@ async function startUpload(t) {{
 
     const need = [];
     for (let i = 0; i < t.totalChunks; i++) if (!t.uploaded.has(i)) need.push(i);
-    await runPool(need, CONCURRENCY, i => uploadChunk(t, i));
+    await runPool(need, CHUNK_CONCURRENCY, i => uploadChunk(t, i));
     if (t.uploaded.size < t.totalChunks) throw new Error("部分分片未能上传");
 
     setUploadState(t, "hashing");
@@ -359,6 +361,9 @@ async function startUpload(t) {{
     meta.className = "task-meta fail-msg";
     meta.textContent = String(e.message || e);
     t.el.appendChild(meta);
+  }} finally {{
+    // 无论成功还是失败，都触发 pump 让队列中下一个文件开始上传
+    pump();
   }}
 }}