Files
xiaoxia-saas/apps/worker/worker_app/celery_app.py
T
CI Bot a33532113b fix(voice-clone): 处理中卡死兜底 — worker重启/Celery消息丢失后processing记录永久卡住问题
根因:
- 声音克隆走 API提交CosyVoice→Celery worker轮询5分钟→mark ready/failed
- 当worker容器重启、进程OOM或Celeryprefetch消息丢失时,已进入processing状态的voice_clone_profile没有兜底恢复机制
- 对比ingest链路有cleanup_stale_ingest_jobs beat任务+worker_ready启动恢复,voice_clone链路缺失
- 结果: 用户看到"克隆处理中,请稍候..."永久转圈,刷新也不变

修复:
1. SQLAlchemyVoiceCloneProfileRepository新增cleanup_stale_processing方法:
   扫描updated_at超过10分钟的processing记录(正常克隆<5分钟),标记failed并提示用户重试
2. _startup.py新增recover_stale_voice_clones_on_startup+worker_ready信号:
   worker启动时一次性扫描恢复,用户重启/部署后立即解锁卡死任务
3. cleanup.py新增scheduled_cleanup_stale_voice_clones beat任务:
   每5分钟巡检兜底,防止运行期OOM/消息丢失导致的新卡死
4. celery_app.py beat_schedule注册新任务
5. voice_clone.py任务入口加received日志,便于日志排查
2026-09-26 14:51:57 +08:00

86 lines
4.3 KiB
Python
Executable File
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
from celery import Celery
from worker_app.core.config import get_settings
settings = get_settings()
celery_app = Celery(settings.worker_name)
celery_app.conf.broker_url = settings.broker_url
celery_app.conf.result_backend = settings.result_backend
celery_app.conf.broker_connection_retry_on_startup = True
# #1714 队列隔离:generation(高优,独占 worker)/ transcode(素材转码)/ celery(默认)
from packages.shared.celery_queues import ( # noqa: E402
GENERATION_WORKER_PREFETCH_MULTIPLIER,
apply_queue_settings,
)
apply_queue_settings(celery_app)
# 长渲染任务预取 1,避免任务被预取占住导致调度不均
celery_app.conf.worker_prefetch_multiplier = GENERATION_WORKER_PREFETCH_MULTIPLIER
celery_app.conf.task_acks_late = True # worker 崩溃时未完成任务重回队列,由执行前守卫丢弃作废消息
# worker 进程被 OOM/容器硬杀时拒绝 ack,消息留在队列由其他 worker 接手
celery_app.conf.task_reject_on_worker_lost = True
# Redis broker 消息可见性超时(#1714):acks_late 下,消息被预取后 visibility_timeout
# 内未 ack 才会重投。长任务(ingest HEVC 转码 20-30 分钟、生成硬超时 11 分钟)
# 必须远大于最长执行时间,否则正常任务会在执行中被误重投;4 小时覆盖最长转码 + 余量。
celery_app.conf.broker_transport_options = {"visibility_timeout": 4 * 60 * 60}
celery_app.conf.imports = (
"worker_app.tasks.health",
"worker_app.tasks.ingest",
"worker_app.tasks.atom_clips",
# #1970 片段级 AI 标签:必须显式 import 注册,否则 worker 报
# "Received unregistered task of type 'worker.tag_atom_clip'"
"worker_app.tasks.atom_clip_tagging",
"worker_app.tasks.asset_quality_scoring_task",
"worker_app.tasks.backfill_atom_clip_tags",
"worker_app.tasks.classification",
"worker_app.tasks.generation",
"worker_app.tasks.voice_extraction",
"worker_app.tasks.voice_clone",
"worker_app.tasks.tts_synthesis",
"worker_app.tasks.batch_download",
"worker_app.tasks.duplication_check",
# #1798 AI 数字人渲染:必须在 Worker 实例上注册同名任务,否则消息无人消费(渲染卡 0%)
"worker_app.tasks.ai_avatar_render",
"worker_app.tasks._startup",
"apps.worker.video_processing.dedup",
"worker_app.tasks.cleanup",
# 注意:必须用 app.* 路径,不能用 apps.api.app.* 路径!
# PYTHONPATH=/app/apps/api 下,app.tasks.lipsync_tts 可直接导入且不触发 apps/api/__init__.py
# (apps/api/__init__.py 会 from .main import app,级联加载整个 FastAPI 栈,Worker 中不需要且会导致注册失败)
"app.tasks.lipsync_tts",
# #1998 GPU MuseTalk 异步推理:wait_for_result→签名 URL→回写 lipsync_jobs
# 必须在 Worker 侧注册,否则 apply_async 消息无人消费,job 永远卡在 processing
"app.tasks.lipsync_gpu",
)
# Celery Beat 定时任务调度
# 注:worker 单实例内嵌 beat(entrypoint-worker.sh -B),定时任务不会重复执行
celery_app.conf.beat_schedule = {
# pending 任务超时清理:worker 停止消费后,卡 pending 的任务 15 分钟内释放限流名额
"cleanup-stale-pending-tasks": {
"task": "worker.cleanup_stale_pending_tasks",
"schedule": 300.0, # 每 5 分钟(秒)
"options": {"expires": 240}, # 4 分钟过期,避免堆积
},
# running 孤儿任务巡检:容器重启/进程被杀后卡 running 的任务,20 分钟无更新则判失败
"cleanup-stale-running-tasks": {
"task": "worker.cleanup_stale_running_tasks",
"schedule": 300.0, # 每 5 分钟(秒)
"options": {"expires": 240},
},
# 上传/转码链路孤儿巡检:worker 重启丢 prefetch 消息后,卡 pending/processing
# 的 ingest_job + asset 占位超时标终态(#1714)。转码任务较长,10 分钟一轮
"cleanup-stale-ingest-jobs": {
"task": "worker.cleanup_stale_ingest_jobs",
"schedule": 600.0, # 每 10 分钟(秒)
"options": {"expires": 540},
},
# 音色克隆卡死巡检:worker 重启/消息丢失后 processing 卡 10 分钟标 failed,用户可点重试
"cleanup-stale-voice-clones": {
"task": "worker.cleanup_stale_voice_clones",
"schedule": 300.0, # 每 5 分钟
"options": {"expires": 240},
},
}