graph LR
    A[媒体网关] -->|SRTP 解密| B(转码/混流引擎)
    B --> C{合规策略引擎}
    C -->|隐形水印嵌入| D[对象存储]
    C -->|明文审计日志| E[合规归档]
    D --> F[水印提取溯源]

// 伪代码：NUMA 本地化内存分配策略
struct gateway_ctx {
    struct rte_mempool *mbuf_pool[NR_NUMA_NODES];
    struct rte_ring *rx_ring[NR_NUMA_NODES];
    struct lcore_conf lcore_conf[RTE_MAX_LCORE];
};

void init_numa_aware() {
    for (int node = 0; node < numa_max_node(); ++node) {
        ctx->mbuf_pool[node] = rte_pktmbuf_pool_create(
            "mbuf_pool", 8192, 256, 0, RTE_MBUF_DEFAULT_BUF_SIZE, node);
        // 绑定 RX 队列到同 NUMA 核心
        rte_eth_dev_rx_queue_setup(port, qid, 1024, node, NULL, ctx->mbuf_pool[node]);
    }
}

// 伪代码：转码任务调度决策树
func SelectTranscoder(task *TranscodeTask) (Worker, error) {
    // 1. 硬件能力标签匹配
    candidates := filterByCodecAndProfile(task.Codec, task.Profile)
    // 2. 负载评分：GPU 显存占用 < 80%、CPU 负载 < 70%、队列深度 < 10
    candidates = filterByLoad(candidates)
    // 3. 亲和性优先：同 NUMA、同可用区、同机架
    return scoreAndPick(candidates, task.AffinityLabels)
}

sequenceDiagram
    participant GW as 媒体网关
    participant AI as AI 推理服务
    participant MCU as 混流/录制
    GW->>AI: RTP 载荷 + 元数据 (SSRC, Codec, TS)
    AI-->>GW: 增强后载荷 / 推理结果 (VAD/ASR/Layout)
    GW->>MCU: 转发处理后流
    Note right of GW: 零拷贝共享内存 / GPU Direct RDMA

# .gitlab-ci.yml 片段：多架构并行构建与测试
stages:
  - build_multiarch
  - unit_test
  - integration_test_arm64
  - integration_test_x86
  - security_scan
  - deploy_staging

build:arm64:
  stage: build_multiarch
  image: docker.io/library/golang:1.22
  script:
    - docker buildx create --use --driver docker-container --platform linux/arm64,linux/amd64
    - docker buildx build --platform linux/arm64 -t $REGISTRY/$IMAGE:arm64-$CI_COMMIT_SHA --push .
  tags: [kylin-v10-arm64]

integration_test_arm64:
  stage: integration_test_arm64
  needs: [build:arm64]
  script:
    - helm upgrade --install gateway ./chart --set image.tag=arm64-$CI_COMMIT_SHA --set arch=arm64
    - pytest tests/integration --junitxml=report.xml
  tags: [kylin-v10-arm64]

# 测试用例自动生成策略（伪代码）
class InteropMatrixGenerator:
    def generate(self):
        # 1. 核心矩阵：主流终端最新 3 版本 + LTS 版本 全排列
        core_matrix = cartesian_product(MAIN_VENDORS, LATEST_3_VERSIONS + LTS)
        # 2. 长尾矩阵：历史版本随机抽样 + 边界参数 (极高/低码率、奇葩 SDP)
        long_tail = sampled_matrix(ALL_VERSIONS, rate=0.1) + boundary_cases()
        # 3. 回归矩阵：近 3 个发布版本修复的 Bug 对应用例
        regression = collect_fixed_bug_cases(last_n_releases=3)
        return deduplicate(core_matrix + long_tail + regression)

graph TD
    A[代码合入] --> B(单元测试/静态扫描)
    B --> C{镜像构建<br/>多架构}
    C --> D[部署临时测试环境<br/>Kind/K3s + 模拟终端]
    D --> E[并行执行互通用例集]
    E --> F{通过率 ≥ 99.5%<br/>无 P0 失败}
    F -->|Yes| G[推送镜像仓库<br/>打标签 rc-xxx]
    F -->|No| H[自动创建缺陷单<br/>阻断合并]
    G --> I[预发环境压测/混沌工程]
    I --> J[生产灰度发布]