
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>HANS</title>
      <link>https://www.hansking.cn/blog</link>
      <description>A posts created with Next.js and Tailwind.css</description>
      <language>en-us</language>
      <managingEditor>admin@hansking.cn (HANS)</managingEditor>
      <webMaster>admin@hansking.cn (HANS)</webMaster>
      <lastBuildDate>Wed, 24 Jun 2026 10:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.hansking.cn/tags/可观测性/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.hansking.cn/blog/2026/AI-xi-tong-de-SLO-zhi-liang-yan-chi-ke-yong-xing-yu-an-quan</guid>
    <title>AI 系统的 SLO：质量、延迟、可用性与安全</title>
    <link>https://www.hansking.cn/blog/2026/AI-xi-tong-de-SLO-zhi-liang-yan-chi-ke-yong-xing-yu-an-quan</link>
    <description>[{&quot;value&quot;:&quot;AI 系统的 SLO：质量、延迟、可用性与安全&quot;,&quot;url&quot;:&quot;#ai-系统的-slo质量延迟可用性与安全&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先分清 SLI、SLO 和 SLA&quot;,&quot;url&quot;:&quot;#一先分清-slislo-和-sla&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、AI 系统至少需要四类 SLO&quot;,&quot;url&quot;:&quot;#二ai-系统至少需要四类-slo&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;1. 质量 SLO&quot;,&quot;url&quot;:&quot;#1-质量-slo&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;2. 延迟 SLO&quot;,&quot;url&quot;:&quot;#2-延迟-slo&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;3. 可用性 SLO&quot;,&quot;url&quot;:&quot;#3-可用性-slo&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;4. 安全 SLO&quot;,&quot;url&quot;:&quot;#4-安全-slo&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;三、质量指标要有采样和标签&quot;,&quot;url&quot;:&quot;#三质量指标要有采样和标签&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、错误预算让团队敢于做改进&quot;,&quot;url&quot;:&quot;#四错误预算让团队敢于做改进&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、降级要提前设计，不要故障时临时发挥&quot;,&quot;url&quot;:&quot;#五降级要提前设计不要故障时临时发挥&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、发布门禁要同时看离线和在线&quot;,&quot;url&quot;:&quot;#六发布门禁要同时看离线和在线&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、告警要连接到行动&quot;,&quot;url&quot;:&quot;#七告警要连接到行动&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、SLO 也要防止被滥用&quot;,&quot;url&quot;:&quot;#八slo-也要防止被滥用&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：SLO 是团队对用户体验的共同承诺&quot;,&quot;url&quot;:&quot;#总结slo-是团队对用户体验的共同承诺&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Wed, 24 Jun 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>SLO</category><category>可观测性</category><category>可靠性</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026/LLM-wang-guan-she-ji-tong-yi-jian-quan-xian-liu-chong-shi-yu-shen-ji</guid>
    <title>LLM 网关设计：统一鉴权、限流、重试与审计</title>
    <link>https://www.hansking.cn/blog/2026/LLM-wang-guan-she-ji-tong-yi-jian-quan-xian-liu-chong-shi-yu-shen-ji</link>
    <description>[{&quot;value&quot;:&quot;LLM 网关设计：统一鉴权、限流、重试与审计&quot;,&quot;url&quot;:&quot;#llm-网关设计统一鉴权限流重试与审计&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、网关到底负责什么&quot;,&quot;url&quot;:&quot;#一网关到底负责什么&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、先定义统一请求协议&quot;,&quot;url&quot;:&quot;#二先定义统一请求协议&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、鉴权和授权要分开&quot;,&quot;url&quot;:&quot;#三鉴权和授权要分开&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、模型路由要可解释&quot;,&quot;url&quot;:&quot;#四模型路由要可解释&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、限流要同时限制请求和资源&quot;,&quot;url&quot;:&quot;#五限流要同时限制请求和资源&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、重试要防止重复副作用&quot;,&quot;url&quot;:&quot;#六重试要防止重复副作用&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、熔断和降级要保留安全边界&quot;,&quot;url&quot;:&quot;#七熔断和降级要保留安全边界&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、审计日志要足够排查，但不要保存一切&quot;,&quot;url&quot;:&quot;#八审计日志要足够排查但不要保存一切&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、成本和质量要从同一条链路统计&quot;,&quot;url&quot;:&quot;#九成本和质量要从同一条链路统计&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：网关不是代理，而是规则的共同入口&quot;,&quot;url&quot;:&quot;#总结网关不是代理而是规则的共同入口&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Tue, 13 Jan 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>API Gateway</category><category>限流</category><category>可观测性</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026/LLM-guan-ce-TracingToken-yong-liang-yu-shi-bai-yang-ben</guid>
    <title>LLM 观测：Tracing、Token 用量与失败样本</title>
    <link>https://www.hansking.cn/blog/2026/LLM-guan-ce-TracingToken-yong-liang-yu-shi-bai-yang-ben</link>
    <description>[{&quot;value&quot;:&quot;LLM 观测：Tracing、Token 用量与失败样本&quot;,&quot;url&quot;:&quot;#llm-观测tracingtoken-用量与失败样本&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先把一次 LLM 请求拆成 Span&quot;,&quot;url&quot;:&quot;#一先把一次-llm-请求拆成-span&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、Trace ID 不是用户 ID&quot;,&quot;url&quot;:&quot;#二trace-id-不是用户-id&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、Token 计量不只是为了算账&quot;,&quot;url&quot;:&quot;#三token-计量不只是为了算账&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、延迟要拆成用户真正等待的阶段&quot;,&quot;url&quot;:&quot;#四延迟要拆成用户真正等待的阶段&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、错误要分成可行动的类别&quot;,&quot;url&quot;:&quot;#五错误要分成可行动的类别&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、失败样本是最有价值的训练材料&quot;,&quot;url&quot;:&quot;#六失败样本是最有价值的训练材料&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、在线质量监控不能只靠用户点赞&quot;,&quot;url&quot;:&quot;#七在线质量监控不能只靠用户点赞&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、告警要能推动动作&quot;,&quot;url&quot;:&quot;#八告警要能推动动作&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、隐私与调试之间要划边界&quot;,&quot;url&quot;:&quot;#九隐私与调试之间要划边界&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十、Trace 回放要避免产生真实副作用&quot;,&quot;url&quot;:&quot;#十trace-回放要避免产生真实副作用&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十一、观测数据本身也要版本化&quot;,&quot;url&quot;:&quot;#十一观测数据本身也要版本化&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十二、把观测接到研发闭环&quot;,&quot;url&quot;:&quot;#十二把观测接到研发闭环&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十三、我的总结：观测不是给模型装摄像头&quot;,&quot;url&quot;:&quot;#十三我的总结观测不是给模型装摄像头&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 10 Nov 2024 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>可观测性</category><category>Tracing</category><category>工程实践</category>
  </item>

    </channel>
  </rss>
