优化架构K8s 副本和队列

.gitea/workflows/deploy.yaml

@@ -147,7 +147,7 @@ jobs:
           sed -i "s|airshelf.airlabs.art|${{ env.DOMAIN_WEB }}|g" k8s/ingress.yaml
 
           # ===== Core (real app) image + domain substitution =====
-          sed -i "s|\${CI_REGISTRY_IMAGE}/airshelf-core-api:latest|${CR_IMAGE}/airshelf-core-api:${{ env.IMAGE_TAG }}|g" k8s/core/api-deployment.yaml k8s/core/worker-deployment.yaml
+          sed -i "s|\${CI_REGISTRY_IMAGE}/airshelf-core-api:latest|${CR_IMAGE}/airshelf-core-api:${{ env.IMAGE_TAG }}|g" k8s/core/api-deployment.yaml k8s/core/worker-deployment.yaml k8s/core/worker-quick-deployment.yaml
           sed -i "s|\${CI_REGISTRY_IMAGE}/airshelf-core-web:latest|${CR_IMAGE}/airshelf-core-web:${{ env.IMAGE_TAG }}|g" k8s/core/web-deployment.yaml
           sed -i "s|airshelf-web.airlabs.art|${{ env.DOMAIN_CORE }}|g" k8s/core/ingress.yaml
 
@@ -196,6 +196,7 @@ jobs:
               # Core real app (api + celery worker + web + ingress)
               kubectl $KUBECTL_TIMEOUT apply -f k8s/core/api-deployment.yaml
               kubectl $KUBECTL_TIMEOUT apply -f k8s/core/worker-deployment.yaml
+              kubectl $KUBECTL_TIMEOUT apply -f k8s/core/worker-quick-deployment.yaml
               kubectl $KUBECTL_TIMEOUT apply -f k8s/core/web-deployment.yaml
               kubectl $KUBECTL_TIMEOUT apply -f k8s/core/ingress.yaml
 
@@ -205,6 +206,7 @@ jobs:
               kubectl $KUBECTL_TIMEOUT rollout restart deployment/airshelf-web
               kubectl $KUBECTL_TIMEOUT rollout restart deployment/airshelf-core-api
               kubectl $KUBECTL_TIMEOUT rollout restart deployment/airshelf-core-worker
+              kubectl $KUBECTL_TIMEOUT rollout restart deployment/airshelf-core-worker-quick
               kubectl $KUBECTL_TIMEOUT rollout restart deployment/airshelf-core-web
             } 2>&1 | tee /tmp/deploy.log && { ok=1; break; }
             echo "Attempt $attempt failed, retrying in 30s..."

core/backend/airshelf/settings/base.py

@@ -204,6 +204,14 @@ CELERY_TASK_ACKS_LATE = True
 CELERY_TASK_REJECT_ON_WORKER_LOST = True
 CELERY_WORKER_PREFETCH_MULTIPLIER = 1
 CELERY_TIMEZONE = TIME_ZONE
+# 快活走 airshelf.quick(独立 worker-quick Pod);出图/出片仍默认 celery 队列。
+CELERY_TASK_ROUTES = {
+    "apps.ai.tasks.extract_entities_task": {"queue": "airshelf.quick"},
+    "apps.ai.tasks.poll_free_video_task": {"queue": "airshelf.quick"},
+    "apps.ai.tasks.drain_asset_reviews_task": {"queue": "airshelf.quick"},
+    "apps.ops.tasks.ensure_team_notifications_task": {"queue": "airshelf.quick"},
+    "apps.projects.tasks.run_quick_script_task": {"queue": "airshelf.quick"},
+}
 
 REDIS_LOCK_URL = env("REDIS_LOCK_URL", "redis://127.0.0.1:6379/3")
 

core/backend/apps/ai/tasks.py

@@ -1,5 +1,7 @@
 from airshelf.celery import app
 
+QUICK_QUEUE = "airshelf.quick"
+
 
 @app.task(bind=True, max_retries=3)
 def submit_ai_task(self, task_id: str) -> str:
@@ -21,7 +23,7 @@ def generate_standalone_image_task(self, task_id: str) -> str:
     return task_id
 
 
-@app.task(bind=True, max_retries=0)
+@app.task(bind=True, max_retries=0, queue=QUICK_QUEUE)
 def extract_entities_task(self, task_id: str) -> str:
     """实体提取的慢活(豆包思考模型流式,可达数十秒)在 worker 内跑,Web 层秒回不被占住、不再 502。
     幂等且失败自退费(见 run_extract_entities_task),故 max_retries=0,不向上抛重试。"""
@@ -83,7 +85,7 @@ def generate_model_triview_task(self, task_id: str) -> str:
     return task_id
 
 
-@app.task(bind=True, max_retries=0)
+@app.task(bind=True, max_retries=0, queue=QUICK_QUEUE)
 def poll_free_video_task(self, task_id: str, attempt: int = 0) -> str:
     """自由创作视频·worker 兜底轮询:每 30s 一次自重排(不依赖 celery beat),
     最多 60 次(约 30 分钟);用尽后只停止 Worker 兜底,不把业务任务判失败。
@@ -119,7 +121,7 @@ def poll_free_video_task(self, task_id: str, attempt: int = 0) -> str:
     return task_id
 
 
-@app.task(bind=True, max_retries=0)
+@app.task(bind=True, max_retries=0, queue=QUICK_QUEUE)
 def drain_asset_reviews_task(self) -> str:
     """全平台人脸审核兜底:未送审自动送火山、审核中自动拉绿/红盾。
     自重排(无 celery beat);eager 下不自转,避免单测死循环。"""

core/backend/apps/ops/tasks.py

@@ -4,7 +4,7 @@ from airshelf.celery import app
 from apps.accounts.models import Team
 
 
-@app.task
+@app.task(queue="airshelf.quick")
 def ensure_team_notifications_task(team_id: str, user_id: str | None = None) -> None:
     """后台生成/补齐团队站内通知。
 

core/frontend/src/App.tsx

@@ -453,7 +453,13 @@ export function App() {
     }
     const active = detail.video_segments.filter((segment) => ["running", "queued"].includes(segment.status));
     if (active.length === 0) return;
-    await Promise.all(active.map((segment) => api.pollVideo(activeProjectId, segment.id).catch(() => undefined)));
+    const polls = await Promise.all(active.map((segment) => api.pollVideo(activeProjectId, segment.id).catch(() => undefined)));
+    // 仍在 queued/running 就别拉整棵项目树(26KB);只有某段终态才回读详情。
+    const settled = polls.some((row) => {
+      const status = row && typeof row === "object" ? (row as { status?: string }).status : "";
+      return Boolean(status) && status !== "running" && status !== "queued";
+    }) || polls.some((row) => row && typeof row === "object" && "id" in (row as object));
+    if (!settled) return;
     const next = await api.project(activeProjectId).catch(() => null);
     // 晚到的旧项目轮询结果不要冲掉已切换的当前项目详情
     if (next && next.id === activeProjectIdRef.current) {

core/frontend/src/routes/pipeline.tsx

@@ -2668,8 +2668,17 @@ export function PipelinePage(props: {
   const activeVideoCount = segments.filter((s) => ["running", "queued"].includes(s.status)).length;
   useEffect(() => {
     if (activeVideoCount === 0) return;
-    const timer = window.setInterval(() => { void onPollVideosQuiet(); }, 5000);
+    const tick = () => {
-    return () => window.clearInterval(timer);
+      if (document.hidden) return;
+      void onPollVideosQuiet();
+    };
+    const timer = window.setInterval(tick, 5000);
+    const onVis = () => { if (!document.hidden) void onPollVideosQuiet(); };
+    document.addEventListener("visibilitychange", onVis);
+    return () => {
+      window.clearInterval(timer);
+      document.removeEventListener("visibilitychange", onVis);
+    };
   }, [activeVideoCount, onPollVideosQuiet]);
 
   // 进入视频 / 拼接阶段时回填已有成片(此前合成过就直接给出播放/下载入口)

k8s/core/api-deployment.yaml

@@ -5,7 +5,7 @@ metadata:
   labels:
     app: airshelf-core-api
 spec:
-  replicas: 1
+  replicas: 2
   selector:
     matchLabels:
       app: airshelf-core-api

k8s/core/worker-deployment.yaml

@@ -23,7 +23,8 @@ spec:
         # Celery worker connects to the external (Volcano managed) Redis broker
         # configured via the airshelf-core-env secret. Uses `args` (not `command`)
         # so the image ENTRYPOINT still runs but skips migrate/collectstatic ($1=celery).
-        args: ["celery", "-A", "airshelf.celery:app", "worker", "-l", "info", "-Q", "celery,airshelf.quick", "--concurrency", "4"]
+        # 本 Pod 只吃默认 celery 队列(出图/出片等慢活)。提取/通知/脚本思考在 worker-quick。
+        args: ["celery", "-A", "airshelf.celery:app", "worker", "-l", "info", "-Q", "celery", "--concurrency", "4"]
         envFrom:
         - secretRef:
             name: airshelf-core-env

k8s/core/worker-quick-deployment.yaml

@@ -0,0 +1,42 @@
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: airshelf-core-worker-quick
+  labels:
+    app: airshelf-core-worker-quick
+spec:
+  replicas: 1
+  selector:
+    matchLabels:
+      app: airshelf-core-worker-quick
+  template:
+    metadata:
+      labels:
+        app: airshelf-core-worker-quick
+    spec:
+      imagePullSecrets:
+      - name: cr-pull-secret
+      containers:
+      - name: airshelf-core-worker-quick
+        image: ${CI_REGISTRY_IMAGE}/airshelf-core-api:latest
+        imagePullPolicy: Always
+        # 只消费 airshelf.quick:提取角色、通知补齐、脚本思考、视频状态轻轮询。
+        # 与出片 worker 隔离,Seedance 占满时提取仍能跑。
+        args: ["celery", "-A", "airshelf.celery:app", "worker", "-l", "info", "-Q", "airshelf.quick", "--concurrency", "4"]
+        envFrom:
+        - secretRef:
+            name: airshelf-core-env
+        livenessProbe:
+          exec:
+            command: ["sh", "-c", "celery -A airshelf.celery:app inspect ping"]
+          initialDelaySeconds: 40
+          periodSeconds: 60
+          timeoutSeconds: 15
+          failureThreshold: 3
+        resources:
+          requests:
+            memory: "256Mi"
+            cpu: "100m"
+          limits:
+            memory: "1Gi"
+            cpu: "500m"