源文件入库 → 元数据解析 → 预处理（统一时间基/关键帧/音频重采样）
    → 主转码（多码率/多分辨率/多编码） → 质检（VMAF/PSNR/时长校验）
    → 封装输出（MP4/HLS/DASH/WebM） → 冷热存储分级 → CDN 预热/索引入库

# Intel QSV (oneVPL) - H.264 主流分发
ffmpeg -init_hw_device qsv=hw -filter_hw_device hw 
  -i input.mp4 -vf 'hwupload=extra_hw_frames=64,format=nv12|qsv,hwdownload,format=nv12' 
  -c:v h264_qsv -preset medium -load_plugins 1 
  -global_quality 23 -look_ahead 1 -look_ahead_depth 40 
  -g 60 -keyint_min 60 -bf 3 -refs 3 
  -c:a aac -b:a 128k -ar 48000 -ac 2 
  -movflags +faststart output_qsv.mp4

# NVIDIA NVENC - HEVC 高压缩比分发
ffmpeg -hwaccel cuda -hwaccel_output_format cuda 
  -i input.mp4 
  -c:v hevc_nvenc -preset p4 -tune hq -rc vbr -cq 26 
  -g 60 -keyint_min 60 -bf 3 -refs 3 
  -spatial_aq 1 -temporal_aq 1 
  -c:a aac -b:a 128k -ar 48000 -ac 2 
  -movflags +faststart output_nvenc.mp4

# 强制固定 GOP=2s（30fps -> 60帧），关键帧对齐，场景切换不插入额外 I 帧
-g 60 -keyint_min 60 -sc_threshold 0 -force_key_frames "expr:gte(t,n_forced*2)"

# VMAF 计算（需 ffmpeg 编译 --enable-libvmaf）
ffmpeg -i distorted.mp4 -i reference.yuv -lavfi 
  "libvmaf=model=version=vmaf_v0.6.1:log_fmt=json:log_path=vmaf.json" -f null -

ffmpeg -hwaccel qsv -i input.mp4 
  -map 0:v -map 0:a 
  -c:v h264_qsv -preset medium -global_quality 24 -g 60 -keyint_min 60 
  -c:a aac -b:a 128k -ar 48000 -ac 2 
  -f tee 
  "[select='v:0,a:0':f=mp4:movflags=+faststart]out_1080p.mp4| 
   [select='v:0,a:0':f=hls:hls_time=6:hls_playlist_type=vod:hls_segment_filename=seg_1080p_%04d.ts]out_1080p.m3u8| 
   [select='v:0,a:0':f=dash:window_size=0:seg_duration=6:target_latency=0]out_1080p.mpd| 
   [select='v:0,a:0':f=webm:cluster_time_limit=1000]out_1080p.webm"

ffmpeg -hwaccel qsv -i input.mp4 
  -filter_complex 
  "[0:v]split=4[v0][v1][v2][v3]; 
   [v0]scale_qsv=w=1920:h=1080[out0]; 
   [v1]scale_qsv=w=1280:h=720[out1]; 
   [v2]scale_qsv=w=854:h=480[out2]; 
   [v3]scale_qsv=w=640:h=360[out3]" 
  -map "[out0]" -map 0:a -c:v:0 h264_qsv -b:v:0 4500k -g 60 -keyint_min 60 
  -map "[out1]" -map 0:a -c:v:1 h264_qsv -b:v:1 2500k -g 60 -keyint_min 60 
  -map "[out2]" -map 0:a -c:v:2 h264_qsv -b:v:2 1200k -g 60 -keyint_min 60 
  -map "[out3]" -map 0:a -c:v:3 h264_qsv -b:v:3 600k  -g 60 -keyint_min 60 
  -c:a aac -b:a 128k -ar 48000 -ac 2 
  -f tee 
  "[select='v:0,a:0':f=hls:hls_time=6:hls_playlist_type=vod:hls_segment_filename=1080p_%04d.ts:master_pl_name=master.m3u8]1080p.m3u8| 
   [select='v:1,a:0':f=hls:hls_time=6:hls_playlist_type=vod:hls_segment_filename=720p_%04d.ts]720p.m3u8| 
   [select='v:2,a:0':f=hls:hls_time=6:hls_playlist_type=vod:hls_segment_filename=480p_%04d.ts]480p.m3u8| 
   [select='v:3,a:0':f=hls:hls_time=6:hls_playlist_type=vod:hls_segment_filename=360p_%04d.ts]360p.m3u8"

ffmpeg -i input.mp4 
  -c:v libvpx-vp9 -b:v 2M -crf 30 -speed 2 -tile-columns 2 -frame-parallel 1 -g 60 -keyint_min 60 
  -c:a libopus -b:a 128k -ar 48000 -ac 2 
  -f webm -cluster_time_limit 1000 output_vp9.webm

# 单轨：降噪 + 响度标准化 (EBU R128 -23 LUFS) + 重采样 48kHz
ffmpeg -i input.mp4 -af 
  "afftdn=nr=12:nf=-25, 
   loudnorm=I=-23:TP=-2:LRA=7:measured_I=-99:measured_TP=-99:measured_LRA=0:measured_thresh=-99:offset=0:linear=true:print_format=json, 
   aresample=48000:resampler=soxr:precision=28:osf=2:cutoff=0.99" 
  -c:v copy -c:a aac -b:a 128k output_audio_fixed.mp4

# 假设源文件含 3 路音频流 0:a:0(主讲) 0:a:1(共享) 0:a:2(系统)
ffmpeg -i input.mp4 
  -map 0:v -map 0:a:0 -c:v copy -c:a aac -b:a 128k speaker.mp4 
  -map 0:v -map 0:a:1 -c:v copy -c:a aac -b:a 128k screenshare.mp4 
  -map 0:v -map 0:a:2 -c:v copy -c:a aac -b:a 128k system.mp4

# 伪代码：任务定义（可序列化存入 Redis/DB）
class TranscodeJob:
    source_uri: str
    source_meta: MediaProbeResult
    profile: str          # "distribution" | "archive" | "abr_hls" | "webm_vp9"
    priority: int         # 0 高 1 正常 2 低
    idempotency_key: str  # source_hash + profile + param_hash
    output_spec: List[OutputTarget]  # 含存储桶、CDN 预热 URL、回调地址

# Dockerfile.ffmpeg-qsv
FROM intel/media-ubuntu:22.04
RUN apt-get update && apt-get install -y --no-install-recommends 
    ffmpeg libvmaf-dev && rm -rf /var/lib/apt/lists/*
# 关键：宿主机需挂载 /dev/dri:/dev/dri 并 --device=/dev/dri/renderD128
ENTRYPOINT ["ffmpeg"]

# k8s Deployment 片段
resources:
  limits:
    intel.com/gpu: "1"          # 或 nvidia.com/gpu: "1"
    cpu: "4"
    memory: "8Gi"
volumeMounts:
- name: dri
  mountPath: /dev/dri
volumes:
- name: dri
  hostPath:
    path: /dev/dri

转码完成事件 (Kafka/EventBridge) 
  → AI 编排器 (DAG: ASR → Speaker Diarization → LLM Summary → Chaptering → Vector Embedding)
  → 结构化写入 (ES/ClickHouse/Milvus) 
  → 回调通知前端/IM 机器人

# 1. 单声道 16kHz PCM（ASR 标准输入），切除静音段降成本
ffmpeg -i input.mp4 -af "aresample=16000:resampler=soxr,pan=mono|c0=c0, silenceremove=start_periods=1:start_threshold=-50dB:start_silence=0.5:stop_periods=-1:stop_threshold=-50dB:stop_silence=0.5" -f wav -acodec pcm_s16le asr_input.wav

# 2. 并行产出“可视化波形图”供前端时间轴定位（极小体积）
ffmpeg -i input.mp4 -filter_complex "[0:a]showwavespic=s=1200x120:colors=0x00BFFF:waveform=1[wav]" -frames:v 1 waveform.png

# 3. 关键帧截图做多模态理解（每 30s 1 张，或场景切换检测）
ffmpeg -i input.mp4 -vf "select='gt(scene,0.3)',scale=640:360,tile=5x5" -vsync vfr -q:v 2 thumbs_%04d.jpg

{
  "meeting_meta": { "title": "Q3 产品规划评审", "date": "2025-08-15", "duration_sec": 3720, "confidentiality": "internal" },
  "speakers": [ {"id": "S01", "name": "张三", "role": "PM", "speak_ratio": 0.35}, {"id": "S02", "name": "李四", "role": "Tech Lead", "speak_ratio": 0.28} ],
  "chapters": [
    {"start": 0, "end": 600, "title": "背景与目标对齐", "summary": "确认 Q3 核心指标为留存与付费转化..."},
    {"start": 600, "end": 1800, "title": "技术方案评审", "summary": "对比方案 A/B，敲定采用方案 B，风险点在于..."}
  ],
  "action_items": [
    {"owner": "李四", "task": "输出方案 B 详细设计文档", "deadline": "2025-08-20", "priority": "P0"},
    {"owner": "王五", "task": "准备压测脚本与基线数据", "deadline": "2025-08-18", "priority": "P1"}
  ],
  "keywords": ["留存", "付费转化", "方案 B", "压测"],
  "sentiment": "constructive",
  "vector_id": "vec_asset_12345_chapter_0"
}

# 伪代码：调用编译为 filter 的水印库 (如 libwatermark.so)
ffmpeg -i input.mp4 -vf "watermark=payload='uid=8888|sid=20250815001|ts=1723700000':strength=0.08:seed=0xC0FFEE" -c:v h264_qsv ... output.mp4

// 伪代码：每 2 秒刷新位置/透明度，含用户工号+观看时间+随机噪点
const watermarkText = `Confidential | ${userId} | ${new Date().toLocaleString()} | ${crypto.randomUUID().slice(0,4)}`;
ctx.fillStyle = 'rgba(255,255,255,0.12)'; ctx.font = '18px PingFang SC';
ctx.fillText(watermarkText, randomX(), randomY()); // 随机位置防遮挡

sequenceDiagram
    Client->>Auth Service: 请求播放 Token (含 asset_id, user_id, device_fp, expiry)
    Auth Service-->>Client: Signed JWT (RS256) + Widevine/PlayReady/FairPlay License URL
    Client->>CDN: 请求 HLS/DASH (携带 Token Query/String)
    CDN->>Edge Auth: 验证 Token 签名/过期/设备指纹/并发数
    Edge Auth-->>CDN: 200 OK / 403 Forbidden
    Client->>DRM License Server: 请求 License (携带 Token + Challenge)
    DRM Server-->>Client: License (含 Content Key, 输出保护策略 HDCP 2.2)

# ScaledObject: 基于 Kafka Lag + 任务队列深度 + GPU 显存水位 多维扩缩容
apiVersion: keda.sh/v1alpha1
kind: ScaledObject
metadata:
  name: ffmpeg-transcoder-scaler
spec:
  scaleTargetRef:
    name: ffmpeg-worker-deployment
  pollingInterval: 15
  cooldownPeriod: 300
  minReplicaCount: 0          # 闲时缩至 0，零成本
  maxReplicaCount: 200
  advanced:
    restoreToOriginalReplicaCount: false
    horizontalPodAutoscalerConfig:
      behavior:
        scaleDown:
          stabilizationWindowSeconds: 180
          policies:
          - type: Percent
            value: 10
            periodSeconds: 60
  triggers:
  - type: kafka
    metadata:
      topic: transcode-jobs
      bootstrapServers: kafka:9092
      consumerGroup: keda-scaler
      lagThreshold: "5"       # 每积压 5 个任务扩 1 副本
  - type: prometheus
    metadata:
      serverAddress: http://prometheus:9090
      metricName: gpu_memory_free_bytes
      query: 'sum(nvidia_gpu_memory_free_bytes{pod=~"ffmpeg-.*"}) by (pod)'
      threshold: "2000000000" # 单卡剩余显存 < 2GB 触发扩容
      authModes: "none"

{
  "$schema": "http://json-schema.org/draft-07/schema#",
  "title": "TranscodePlaybackContract",
  "type": "object",
  "required": ["version", "profiles", "drm", "timedMetadata"],
  "properties": {
    "version": { "const": "2.3.0" },
    "profiles": {
      "type": "array",
      "items": {
        "type": "object",
        "required": ["id", "codec", "width", "height", "bitrate", "fps", "gop", "keyint_min", "profile", "level", "audio"],
        "properties": {
          "id": { "enum": ["1080p", "720p", "480p", "360p", "audio_only"] },
          "codec": { "enum": ["avc1.640028", "hev1.1.6.L93.B0", "vp09.00.51.08.01.01.01.01", "av01.0.05M.08"] },
          "gop": { "const": 60 },
          "keyint_min": { "const": 60 },
          "audio": { "codec": "mp4a.40.2", "samplerate": 48000, "channels": 2 }
        }
      }
    },
    "drm": { "systems": ["widevine", "playready", "fairplay"], "keyRotationSec": 21600 },
    "timedMetadata": { "id3": true, "emsg": true, "schemeIdUri": "urn:company:meeting:chapter" }
  }
}

转码端：-g 30 -keyint_min 30 -hls_flags +independent_segments+program_date_time -hls_segment_type fmp4 -target_latency 2
CDN：开启 Chunked Transfer Encoding + HTTP/2 Push
播放端：hls.js + LowLatencyMode (fetch + MediaSource Extensions) 或 原生 CMAF-CTE (Safari/Edge)

// 利用 WebCodecs 硬解 + WASM (ffmpeg.wasm / libvpx.js) 实现帧级拖拽预览
const decoder = new VideoDecoder({
  output: frame => { /* 渲染到 OffscreenCanvas */ },
  error: e => console.error(e)
});
decoder.configure({ codec: 'avc1.640028', codedWidth: 1920, codedHeight: 1080 });
// 仅请求目标关键帧所在段 + 前一段，解码至目标 PTS，生成缩略图/预览帧

主区域 (AZ1+AZ2)          备区域 (AZ3)
   │                         │
   ├── Kafka MirrorMaker ───►│ (异步复制，RPO < 30s)
   ├── 对象存储 CRR ────────►│ (版本控制 + 法律合规锁)
   └── 元数据 DB (TiDB/PolarDB) 同步复制 (RPO=0, RTO<30s)

#!/bin/bash
# check_transcode_env.sh
set -euo pipefail
echo "=== FFmpeg 版本与编码器 ==="
ffmpeg -hide_banner -version | head -1
ffmpeg -hide_banner -encoders | grep -E "qsv|nvenc|amf|libx264|libx265|libvpx-vp9|libopus"

echo -e "n=== 硬件加速设备 ==="
ls -l /dev/dri/ 2>/dev/null || echo "No /dev/dri"
nvidia-smi -L 2>/dev/null || echo "No NVIDIA GPU"
clinfo 2>/dev/null | grep -A5 "Platform Name" || echo "No OpenCL"

echo -e "n=== 编码器功能冒烟测试 (1s testsrc) ==="
for enc in h264_qsv hevc_qsv h264_nvenc hevc_nvenc libx264 libx265; do
  if ffmpeg -hide_banner -encoders | grep -q " $enc "; then
    timeout 10 ffmpeg -v error -f lavfi -i testsrc=duration=1:size=320x240:rate=30 -c:v "$enc" -f null - 2>&1 | head -1
    echo "  [$enc] OK"
  else
    echo "  [$enc] NOT AVAILABLE"
  fi
done

echo -e "n=== VMAF 库 ==="
ffmpeg -v error -i /dev/zero -f lavfi -i testsrc -lavfi libvmaf -f null - 2>&1 | grep -c "VMAF" && echo "VMAF OK" || echo "VMAF MISSING"

# 1. 生产环境抓包 (仅元数据/关键帧，脱敏)
tcpdump -i eth0 -s 0 -w prod_traffic.pcap "port 80 or 443 or 1935" -G 3600 -W 24

# 2. 回放至预发转码集群入口 (Nginx/Envoy)
tcpreplay --intf1=eth0 --loop=0 --mbps=500 --unique-ip prod_traffic.pcap

# 3. 并发模拟器 (Locust 脚本片段)
class TranscodeUser(HttpUser):
    wait_time = between(0.5, 2)
    @task
    def submit_job(self):
        files = {'file': ('meeting_1080p.mp4', open('samples/meeting_1080p.mp4', 'rb'), 'video/mp4')}
        data = {'profile': 'abr_hls', 'priority': 1}
        self.client.post("/api/v1/transcode", files=files, data=data, name="Submit Transcode")