跳到主要内容Skip to main content

运维上线路线The Operator's Route

约 60 分钟about 60 minutes

沿运行、监控、权限、故障与退役读十四个主题。先建立排查顺序,再用具体服务的配置与日志验证。Study fourteen topics covering runtime, monitoring, permissions, failure and retirement. Form an investigation order, then validate it against a service’s configuration and logs.

学完自测Checkpoint
  1. 说出首字延迟和吞吐各自受什么影响,各给一个优化手段
  2. 给线上服务列出三件必须监控的事,并说明各自先查什么
  3. 供应商宣布模型退役时,能走一遍迁移评估的四个检查项
  1. Say what drives time-to-first-token and throughput respectively, with one optimization for each
  2. List three things a live service must monitor, and what to check first for each
  3. When a vendor announces retirement, walk through the four migration-assessment checks

先试着回答,再展开参考答案。答不上来的地方,可从上方步骤返回对应词条。参考回答是核对要点,不要求逐字复述。Try answering before opening the reference. Revisit the relevant step above if needed. These answers highlight key ideas; you do not need to repeat the wording.

第 1 题 · 参考回答Question 1 · Reference answer

首字延迟影响开始等待的时间,吞吐影响连续生成或并发能力。前者可先查输入处理与排队,后者可查批处理与硬件瓶颈;需用相同负载验证。Time to first token affects initial waiting; throughput affects generation/concurrency capacity. Investigate input processing and queues for the former, batching and hardware limits for the latter under equal load.

回看这一题的概念Revisit this concept

第 2 题 · 参考回答Question 2 · Reference answer

先看任务完成率、延迟分布、错误、用量和权限异常。成本突增时对照请求数、输入长度、重试和工具调用记录,不立即归因于模型单价。Track completion, latency distribution, errors, usage and permission anomalies. For a cost spike, compare request count, input lengths, retries and tool calls before blaming the model rate.

回看这一题的概念Revisit this concept

第 3 题 · 参考回答Question 3 · Reference answer

至少覆盖接口与流式、任务质量、工具与权限、成本与延迟,并核对数据条款和回滚。先在代表性请求上对比,达到预先约定的条件再切换。Cover interface/streaming, task quality, tools/permissions and cost/latency, plus data terms and rollback. Compare representative requests before switching against agreed criteria.

回看这一题的概念Revisit this concept

用在一个具体任务里Apply this to a concrete task

从容量预算走到运行检查Move from a capacity budget to runtime checks →