修复线上视频导出卡死(OOM)+ 导出健壮性加固

根因:后端容器内存 limit 仅 768Mi,ffmpeg 编码 1080P 竖屏时内存峰值超限被
K8s OOMKill,导出后台线程随 Pod 重启而死,任务永远停在 35%(本地无 limit 29s 出片)。

- k8s/core/api-deployment.yaml: 后端内存 limit 768Mi→2Gi、requests 256→512Mi,
  给 ffmpeg 足够编码空间(根治)
- export.py: ffmpeg subprocess 加 900s 超时(卡死不再永久 RUNNING)+ -threads 2 降内存峰值
- views.py: submit-export 清理超 5 分钟未结束的僵尸任务,前端不再无限转圈

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

core/backend/apps/projects/services/export.py

@@ -268,7 +268,9 @@ def _build_export_command(*, n: int, specs: list[dict], starts: list[float], tot
     cmd += ["-filter_complex", ";".join(parts), "-map", f"[{vlabel}]"]
     if audio_label:
         cmd += ["-map", f"[{audio_label}]"]
-    cmd += ["-c:v", "libx264", "-pix_fmt", "yuv420p", "-r", fps_expr, "-preset", "veryfast"]
+    # -threads 2:限制 x264 线程数,降低小容器(K8s 内存 limit 小)里的内存峰值,缓解 OOM 被杀;
+    # 同时避免在受限 CPU 上 fork 过多线程争抢。veryfast 保持编码速度。
+    cmd += ["-c:v", "libx264", "-pix_fmt", "yuv420p", "-r", fps_expr, "-preset", "veryfast", "-threads", "2"]
     if audio_label:
         cmd += ["-c:a", "aac", "-b:a", "192k"]
     cmd += ["-t", f"{total:.3f}", "-movflags", "+faststart", "output.mp4"]
@@ -479,7 +481,12 @@ def run_export_job(export_job_id: str) -> ExportJob:
             sub_overlays=sub_overlays, bgm_name=bgm_name, bgm_volume=(bgm_track.volume / 100.0) if bgm_track else 1.0,
             has_audio=has_audio, fps=output_fps, voice_overlays=voice_overlays,
         )
-        proc = subprocess.run(command, cwd=str(tmp), capture_output=True)
+        # 加超时:ffmpeg 卡死(资源不足/被 OOM 杀)时不能让任务永久停在 RUNNING、前端无限转圈;
+        # 超时即按失败收尾,前端能看到「导出失败」而非一直卡。15 分钟足够正常 60s 成片。
+        try:
+            proc = subprocess.run(command, cwd=str(tmp), capture_output=True, timeout=900)
+        except subprocess.TimeoutExpired:
+            raise RuntimeError("ffmpeg 导出超时(15 分钟未完成),通常是服务器资源不足或视频过大,请重试或联系运维扩容")
         if proc.returncode != 0:
             raise RuntimeError(f"ffmpeg export failed: {proc.stderr.decode('utf-8', 'ignore')[-1200:]}")
         output_path = tmp / "output.mp4"

core/backend/apps/projects/views.py

@@ -371,6 +371,15 @@ class ProjectViewSet(TeamScopedViewSetMixin, ModelViewSet):
                         duration_ms=segment.target_duration_seconds * 1000,
                     )
                     start_ms += segment.target_duration_seconds * 1000
+        # 清理僵尸任务:后台线程跑导出时若 Pod 被重启/OOM 杀掉,旧任务会永远停在 RUNNING、
+        # 前端无限轮询。新建前把本时间线超过 5 分钟仍未结束的旧任务标失败,避免界面一直转圈。
+        from django.utils import timezone
+        from datetime import timedelta
+        ExportJob.objects.filter(
+            timeline=timeline,
+            status__in=[ExportJob.Status.QUEUED, ExportJob.Status.RUNNING],
+            updated_at__lt=timezone.now() - timedelta(minutes=5),
+        ).update(status=ExportJob.Status.FAILED, error_message="任务中断(服务重启或超时),已自动失败,请重新导出")
         export_job = create_export_job(timeline=timeline, user=request.user)
         stage, _ = ProjectStage.objects.get_or_create(project=project, stage=ProjectStage.Stage.EXPORT)
         stage.status = ProjectStage.Status.RUNNING

k8s/core/api-deployment.yaml

@@ -51,11 +51,13 @@ spec:
           failureThreshold: 3
         resources:
           requests:
-            memory: "256Mi"
-            cpu: "100m"
+            memory: "512Mi"
+            cpu: "250m"
           limits:
-            memory: "768Mi"
-            cpu: "1000m"
+            # 视频导出在本容器后台线程跑 ffmpeg,编码 1080P 竖屏 + 多路输入 + 字幕/配音混音
+            # 内存峰值可达 1.5GB,原 768Mi 必被 OOMKill(导出卡 35%、Pod 重启)。提到 2Gi 给足空间。
+            memory: "2Gi"
+            cpu: "1500m"
 ---
 apiVersion: v1
 kind: Service