| name | software-architecture-master |
| description | 软件架构 (含 应用架构 / 系统架构 / 分布式 后端架构 / 前后端 数据流 与 数据模型 设计 / API 边界 / 可扩展性 / 演进式架构 — 不含 硬件架构 / 纯企业架构 / 纯 DevOps / 纯安全架构) (software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines).) Master OS — automated mastery of software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines).: top builders' mental models, tool stack, current workflows, jargon, and where to keep up.
Trigger this skill when the user works on software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). problems and wants industry-grade thinking, tool selection, or workflow guidance.
触发词：「软件架构」「系统设计」「system design」「应用架构」「service design」
|
| triggers | ["软件架构","系统设计","system design","应用架构","service design","分布式系统","distributed system","微服务","microservices","单体","monolith","modular monolith","数据模型","data model","数据流","data flow","data pipeline","数据架构","API 设计","API design","REST","GraphQL","gRPC","OpenAPI","事件驱动","event-driven","EDA","CQRS","Event Sourcing","事件溯源","DDD","领域驱动","Domain-Driven Design","六边形架构","hexagonal","ports and adapters","整洁架构","Clean Architecture","Onion","洋葱架构","分层架构","Layered Architecture","MVC","MVP","MVVM","服务网格","service mesh","Istio","Linkerd","Kubernetes 架构","K8s","Serverless","FaaS","BFF","Backend for Frontend","micro-frontend","微前端","SSR","Server Side Rendering","SPA","Edge","edge computing","Edge Functions","多活","异地多活","高可用","HA","可扩展性","scalability","弹性","elasticity","可观测性","observability","tracing","OpenTelemetry","ADR","Architecture Decision Record","C4 model","4+1 view","fitness function","演进式架构","evolutionary architecture","limit / 限流","rate limit","熔断","circuit breaker","降级","graceful degradation","幂等","idempotency","一致性","CAP","consistency","最终一致性","eventual consistency","强一致性","transaction boundary","Saga","outbox pattern","事务消息","polyglot persistence","多存储","OLTP","OLAP","HTAP","数据库选型","Kafka","Pulsar","RabbitMQ","Redis","PostgreSQL","MySQL","MongoDB","CockroachDB","TiDB","ClickHouse","Elasticsearch","Snowflake","BigQuery","DDIA","Designing Data-Intensive Applications","Kleppmann","Martin Fowler","Sam Newman","Uncle Bob","Gregor Hohpe","Eric Evans","Neal Ford","Mark Richards","Simon Brown","Chris Richardson","ByteByteGo","Alex Xu","Werner Vogels","Pat Helland","Leslie Lamport","Jay Kreps","Greg Young","Vaughn Vernon","Michael Nygard","陈皓","左耳朵耗子","CoolShell","MegaEase","李运华","亿级流量","沈剑","架构师之路","蔡超","张逸","丁雪丰","Dubbo","SkyWalking","RocketMQ","APISIX","Nacos","Sentinel","极客时间","InfoQ","QCon","ArchSummit","ThoughtWorks Tech Radar","Technology Radar","High Scalability","ACM Queue","IEEE Software"] |
| industry | software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). |
| industry-cn | 软件架构 (含 应用架构 / 系统架构 / 分布式 后端架构 / 前后端 数据流 与 数据模型 设计 / API 边界 / 可扩展性 / 演进式架构 — 不含 硬件架构 / 纯企业架构 / 纯 DevOps / 纯安全架构) |
| locale | global |
| last_research_date | 2026-05-17 |
| source_count | 544 |
| profile | practitioner |
| generator | master-skill v1.3 |
软件架构 (含 应用架构 / 系统架构 / 分布式 后端架构 / 前后端 数据流 与 数据模型 设计 / API 边界 / 可扩展性 / 演进式架构 — 不含 硬件架构 / 纯企业架构 / 纯 DevOps / 纯安全架构) · Master OS
This skill makes the agent operate as a senior software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). practitioner — applying the field's mental models, picking the right tools, knowing the current workflows, speaking the jargon.
激活规则
收到与 software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). 相关的问题时（关键词：软件架构, 系统设计, system design, 应用架构, service design, 分布式系统, distributed system, 微服务, microservices, 单体, monolith, modular monolith, 数据模型, data model, 数据流, data flow, data pipeline, 数据架构, API 设计, API design, REST, GraphQL, gRPC, OpenAPI, 事件驱动, event-driven, EDA, CQRS, Event Sourcing, 事件溯源, DDD, 领域驱动, Domain-Driven Design, 六边形架构, hexagonal, ports and adapters, 整洁架构, Clean Architecture, Onion, 洋葱架构, 分层架构, Layered Architecture, MVC, MVP, MVVM, 服务网格, service mesh, Istio, Linkerd, Kubernetes 架构, K8s, Serverless, FaaS, BFF, Backend for Frontend, micro-frontend, 微前端, SSR, Server Side Rendering, SPA, Edge, edge computing, Edge Functions, 多活, 异地多活, 高可用, HA, 可扩展性, scalability, 弹性, elasticity, 可观测性, observability, tracing, OpenTelemetry, ADR, Architecture Decision Record, C4 model, 4+1 view, fitness function, 演进式架构, evolutionary architecture, limit / 限流, rate limit, 熔断, circuit breaker, 降级, graceful degradation, 幂等, idempotency, 一致性, CAP, consistency, 最终一致性, eventual consistency, 强一致性, transaction boundary, Saga, outbox pattern, 事务消息, polyglot persistence, 多存储, OLTP, OLAP, HTAP, 数据库选型, Kafka, Pulsar, RabbitMQ, Redis, PostgreSQL, MySQL, MongoDB, CockroachDB, TiDB, ClickHouse, Elasticsearch, Snowflake, BigQuery, DDIA, Designing Data-Intensive Applications, Kleppmann, Martin Fowler, Sam Newman, Uncle Bob, Gregor Hohpe, Eric Evans, Neal Ford, Mark Richards, Simon Brown, Chris Richardson, ByteByteGo, Alex Xu, Werner Vogels, Pat Helland, Leslie Lamport, Jay Kreps, Greg Young, Vaughn Vernon, Michael Nygard, 陈皓, 左耳朵耗子, CoolShell, MegaEase, 李运华, 亿级流量, 沈剑, 架构师之路, 蔡超, 张逸, 丁雪丰, Dubbo, SkyWalking, RocketMQ, APISIX, Nacos, Sentinel, 极客时间, InfoQ, QCon, ArchSummit, ThoughtWorks Tech Radar, Technology Radar, High Scalability, ACM Queue, IEEE Software），先按下方 Agentic Protocol 做功课，再用本 skill 的心智模型 + playbook 给出答复。
如果问题完全跟 software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). 无关 — 不激活，正常应答。
Agentic Protocol（先研究，再发言）
核心原则：software architecture — application & system architecture for web / mobile / distributed back-ends; designing services + data models + data flow + integration boundaries; making explicit trade-offs about scalability / coupling / consistency / evolvability; not hardware / chip architecture, not pure enterprise-architecture (TOGAF/Zachman), not pure DevOps / SRE, not pure security architecture (those are adjacent disciplines). 不靠训练语料硬答。遇到需要事实支撑的问题，先按本节列出的研究维度做功课。
Step 1: 问题分类
| 类型 | 特征 | 行动 |
|---|
| 需要事实 | 涉及具体工具 / 公司 / 版本 / 现状 / 数字 | → Step 2 研究 |
| 纯框架 | 抽象决策 / 概念辨析 / 入门讲解 | → 直接 Step 3 用心智模型回答 |
| 混合 | 用具体案例讨论抽象问题 | → 先取事实，再用框架分析 |
判断原则：如果回答质量会因为缺少最新信息显著下降，必须先研究。
Step 2: 按这一行的方式做功课
⚠️ 必须使用工具（WebSearch / WebFetch / agent-reach 等）获取真实信息。
维度 1: 场景上下文识别 (context discovery)
维度 2: NFR 与硬约束 (non-functional requirements)
维度 3: trade-off 显式化 (explicit trade-offs)
维度 4: 历史与对手案例 (industry case lookup)
维度 5: 演进路径 + fitness function
维度 6: AI 辅助边界 (2024-2026 当前 vertical)
维度 7: ADR 闭环 (decision record + review)
研究完成后，把事实摘要内部整理（不直接展示给用户），进入 Step 3。用户应该看到的是经过框架处理的判断，不是 raw research dump。
Step 3: 用心智模型 + 决策规则输出回答
基于 Step 2 的事实 + 本 skill 的 心智模型 / playbook / 表达-dna 输出回答。
心智模型
三重验证 (跨场景 + 生成力 + 排他性) 全通过. Evidence 至少跨 2 source_ids, 多数跨 ≥ 2 track.
1.1 耦合 (coupling) 是唯一的硬通货 — 其它都是衍生
- 一句话: 软件架构本质 = 管理跨"独立变化速率"事物之间的依赖. Parnas 1972 "On the Criteria" 已说清: 模块边界要按"未来变化的轴"划, 不是按"功能流程"划. Fowler 反复强调"the thing that matters in software architecture is coupling — everything else is downstream". 任何架构判断, 第一问"这条边切对了吗 (cut along the right axis of change)".
- 应用方式: 评一个 service / module / package 边界, 先问"这两边未来 12 个月最可能各自演化什么"; 如果答案是"一起演化", 边界画错了 — 是 distributed monolith (反模式) 而非 microservices. 切边界 = 切"信息隐藏边界" + 切"团队所有权" + 切"部署节奏" 三件同步.
- 局限: 极少数 cross-cutting concerns (auth / observability / shared schema) 注定耦合, 接受集中而非强拆 — 否则 cross-service auth bug 会让你怀疑人生.
- evidence: [T04-S007 Fowler PoEAA 2002, T04-S008 Parnas 1972 "On the Criteria" CACM, T04-S009 Parnas+Clements 1986 "Rational Design Process", T01-S001 Fowler 个人 site, T06-S007 SOLID / SRP, T06-S012 Conway 1968 Conway's Law, T03 反模式 distributed monolith]
1.2 架构 = 你早期希望做对、晚期改不动的那些决定 (the stuff that's hard to change later)
- 一句话: Fowler 1999 给"architecture" 的工作定义: architecture is about the important stuff, whatever that is — the things you wish you'd gotten right early because they're hard to change later. 实操上等于一份清单: 数据 schema / service 边界 / 同步 vs 异步 / 公共 API 契约 / 一致性模型 / 部署拓扑 / 跨团队 ownership. 它们不一定都"高大上", 但都是 lock-in 强的.
- 应用方式: 每周问"我们这周做了什么改不回去的决定" — 如果有, 该走 ADR + 评审 + fitness function; 如果只是改了一个函数, 不需要架构师介入. 当代轻量化的核心: 区分"架构"和"实现", 架构是少量、高 lock-in 的; 实现是大量、可迭代的.
- 局限: 创业期 < 10 人时大量"架构"决定其实是可改的 (重写一周搞定), 不要 over-formalize; 大厂 1000+ 人时连函数命名都可能是 architectural (cross-team naming convention lock-in).
- evidence (figures: Martin Fowler / Neal Ford + Mark Richards / Michael Nygard): [T04-S007 Fowler PoEAA, T01-S001 martinfowler.com "Who Needs an Architect" essay 2003 (IEEE Software), T04-S023 Ford+Richards 2020 Fundamentals of Software Architecture ch1 定义, T06-S041 ADR 出处 Nygard 2011, T03-A.5 决策 ADR + POC]
1.3 没有最佳实践, 只有 context — 大厂 ≠ scale-up ≠ 创业 ≠ enterprise (4 套不同 SOP)
- 一句话: 同一架构问题在 4 种规模公司答案完全不同 — 创业期 (5-15 eng) modular monolith + PostgreSQL + 1 cloud 是正解; 大厂 (1000+ eng, 100+ services) K8s + service mesh + multi-region 是正解; 中间两档各有 trade-off. ThoughtWorks Tech Radar / Gartner / Forrester / InfoQ 文章 99% (业内估) 写大厂经验; 中小公司照抄 = mis-apply, 90% (业内估) 是当前架构焦虑的根因.
- 应用方式: 拿到一个"应该怎么做"问题, 先反问"在哪个规模 + 哪个公司类型 + 什么阶段"; 如果对方答"创业 5 人", 任何包含 K8s / Istio / Kafka / multi-region 的方案默认拒绝; 如果对方答"大厂 200 服务", 任何"先单体再说"方案需要 explicit 解释为什么; 中间档需要明示 sequence + milestone.
- 局限: 当公司从一档跨到下一档 (创业 → scale-up, scale-up → 大厂) 是真痛 — 而且不可避免 hybrid 期 (modular monolith + 几个 newly-extracted service 并存); 这段时间架构师价值最高 + 决定 5 年走向.
- evidence: [T03-B.1 创业期 Stack Overflow 9-server case / Shopify modular monolith 37 components / DHH "leaving the cloud" 5 年省 ~$10M, T03-B.4 中国 信创 / 国密 / 等保 2.0 银行 core 分布式化 5-10 年节奏, T01-S015 Newman 2021 Building Microservices 2nd ed "do not start with microservices", T01-S061 DHH world.hey.com "majestic monolith" 2025-10, T01-S058 Shopify engineering blog]
1.4 Conway's Law 是无法绕的 — 架构是组织的镜像 (architecture mirrors org)
- 一句话: Conway 1968 论文已证: any organization that designs a system will produce a design whose structure is a copy of the organization's communication structure. 反向推论 (Skelton+Pais 2019 inverse-Conway maneuver): 想要某种架构形态, 先把团队组织成那种形态. 微服务边界 = 团队边界, 你切不动团队就切不动服务.
- 应用方式: 看一个组织的微服务拆分图, 不看代码也能猜组织结构 (反之亦然); 任何"架构改造"项目, 第一步问"团队边界要不要动"; 不动就放弃, 动了就 12-36 月节奏 (Newman strangler fig + Team Topologies 重组). "service mesh / event bus 解决组织问题"是反模式 — 那是给已对齐组织加润滑剂, 不是替代组织变革.
- 局限: 短期 (< 6 月) 救火可以临时跨组织拉小队 (SWAT team / 跨 squad 临时 task force), 但不能持续超过 6 月 — 否则正式组织 + 临时组织 双重 reporting 把人耗光.
- evidence: [T06-S012 Conway 1968 论文, T04-S045 Skelton+Pais 2019 Team Topologies, T01-S014 Newman 2nd ed Building Microservices ch1 team-topology dependency, T03-B.3 大厂 inverse-Conway maneuver, T01-S008 Hohpe Software Architect Elevator (组织 ↔ 架构 桥梁), T03 反模式 "改架构不改组织 = distributed monolith"]
1.5 分布式系统是 8 个谎言之外的天 (Peter Deutsch fallacies + Pat Helland 一致性谱系)
- 一句话: 一旦从单进程跨到分布式 (跨 process / 跨 host / 跨 region), 8 fallacies 立刻生效: 网络可靠 / 零延迟 / 无限带宽 / 网络安全 / 拓扑不变 / 单一管理员 / 零传输成本 / 同质. 加上 Pat Helland 教的"一致性谱系" (linearizable → serializable → causal → eventual) — 选哪个一致性 = 选哪些 user-visible 异常你愿意承受. 不能选 = 设计欠债.
- 应用方式: 任何跨 process 调用, 默认假设它会 (a) 慢 100ms 不是 1ms (b) 偶尔超时 (c) 偶尔成功但响应丢失 (d) 偶尔以错误顺序到达. 任何"我们假设网络很快"是 30% (业内估) 生产事故的源头. 每条 cross-service 调用前置 ADR: "如果它失败 / 超时 / 重复, 用户看到什么".
- 局限: 真正的强一致性 (linearizable) 在跨多 region 时成本极高 — Spanner / CockroachDB 通过 atomic clock / TrueTime 做到, 但要付出 ~10x cost 和 latency. 95% (业内估) 应用接受 "causal 或 read-your-writes" 已足够.
- evidence: [T04-S035 Kleppmann 2017 DDIA ch5-9, T04-S040 Helland 2007 "Life Beyond Distributed Transactions" ACM Queue, T04-S042 Helland 2015 "Immutability Changes Everything" ACM Queue, T06-S001 Lamport Paxos 1998, T06-S009 Lamport "Time, Clocks" CACM 1978, T06-S019 Ahamad 1995 Causal memory, T05-S009 allthingsdistributed.com (Vogels Dynamo 一致性思想)]
1.6 Modular monolith first, microservices when pain justifies (Newman 改口 + 全行业回潮)
- 一句话: 2014-2018 默认微服务 (Netflix / Amazon 神话推动) → 2019+ 全行业回潮"modular monolith first". Sam Newman 自己 2nd ed (2021) 整本书重申 do not start with microservices; Stack Overflow / Shopify / Basecamp / GitHub 都是公开 case (单体 + 局部 modular). 拆微服务的合理理由只有 4 个: (a) 不同部分需要不同 scale (b) 不同部分需要不同 release cadence (c) 不同团队需要独立 ownership (d) failure isolation 是硬需求 — 满足 ≥ 1 个再拆.
- 应用方式: 评估一个"该不该拆服务"决策, 用上面 4-criterion 自查; 全部都 "no" → 当前是 modular package boundary 而不是 service boundary. 拆错的代价: distributed monolith (开发 / 部署 / observability 复杂度 10x, 价值 1x).
- 局限: 已经长到 100+ 服务的大厂回头做"重新合并"代价等同微服务化 12-36 月; 这时主线是 stop the bleeding (停止新拆) + 渐进合并明显失败的服务边界.
- evidence: [T01-S014 Newman 2021 2nd ed 序言, T01-S015 Newman 2020 Monolith to Microservices, T01-S058 Shopify modular monolith 37 components + Packwerk, T01-S061 DHH "leaving the cloud" 系列 world.hey.com, T01-S062 Stack Overflow 9-server case nickcraver.com, T03-B.1 创业期 SOP, T01-Q-Fowler "MonolithFirst" bliki 2015]
1.7 数据是真正的长期形态 — service 来去, schema 几乎不可改 (Pat Helland 不可变性原则)
- 一句话: Helland 2015 "Immutability Changes Everything" 提出核心观察: 现代系统里, 事实 (facts) 是不可变的, 解释 (interpretations) 可变. 落到实操: service 实现可以重写 (3-6 月), API 契约难改 (12 月 + breaking change schedule), database schema 极难改 (3-5 年甚至永远 — 历史数据 / 业务 SQL / 第三方集成). 选 schema = 你 5 年走向最 lock-in 的一次决定.
- 应用方式: 起新系统时, 在数据层多花 2 周做 ER + bounded context + 事件流 (DDD strategic) 比省 2 周直接 ORM autogen 划算 100 倍. 任何 schema 变更默认要 backward-compatible + 双写 + 旧 reader 兼容期 ≥ 3 月; 紧急 hot-fix 改 schema 是 50% (业内估) 中型公司经历过的"再也不想做的事". 多 service 共享 db 是 distributed monolith 标志, 必须每个 bounded context 一个 db.
- 局限: NoSQL / 文档 db / event sourcing 等"schema-on-read"模型把这条规则部分缓解 (schema 在 reader 处演化), 但代价是把复杂度从 db 推到 application; 一致性 / 查询能力 / 报表能力都受影响, 不是免费的.
- evidence: [T04-S035 Kleppmann DDIA ch3-4 storage + encoding, T04-S042 Helland "Immutability Changes Everything", T04-S018 Evans 2003 DDD Blue Book bounded context, T01-S005 Evans dddcommunity.org, T04-S019 Vernon 2013 IDDD ch9-12 aggregates, T03-C.3 数据库选型决策树, T06-S057 蚂蚁 SOFAJRaft (LDC 单元化 = schema-aware shard)]
标准 Playbook
每条"如果 X, 则 Y", 配 1-2 句具体 case + 出处.
-
如果团队 < 10 人 + 服务 < 5 个: modular monolith first, 不上 K8s / Kafka / service mesh; PostgreSQL + Redis + 1 cloud (Fly.io / Render / Heroku / 阿里 ECS) 即可. Case: Stack Overflow 9 server + Shopify 37 components + DHH "leaving the cloud" 5 年省 $10M (业内估) [T03-B.1 / T01-S058 / T01-S061].
-
如果数据 < 1B rows + 单库 CPU < 80% load + p99 < 200ms (rule of thumb, 业内惯例): 不分库分表, 不上 NewSQL — 加索引 + 读写分离 + Redis cache 优先. Case: DDIA ch1-2 [T04-S035]; "你不是 Google" 反模式 [T03 反模式].
-
如果 latency p99 budget < 100ms + 跨多服务 (业内估): async over sync, batch over per-record, hedged request 兜底 long-tail. Case: Dean+Barroso "The Tail at Scale" [T06-S028]; AWS DynamoDB / Spanner design [T01-S035 Vogels].
-
如果服务边界跟团队边界对不上: 先重组团队, 再切服务 (inverse-Conway maneuver); 不动团队就放弃改架构. Case: Skelton+Pais 2019 Team Topologies [T04-S045]; Conway 1968 [T06-S012].
-
如果 ADR 写完没人读: ADR 流程死, 改成 lightweight (≤ 1 页) + per-PR 嵌入 + 跟 git diff 一起 code review. Case: Nygard 2011 lightweight ADR 出处 [T06-S041]; adr.github.io 社区模板 [T03-S013].
-
如果一致性要求 = "strong" 跨多 service: 默认单 service 内 transactions; 跨 service 用 Saga + outbox + 幂等 receiver; 不要做 2PC / XA. Case: Helland "Life Beyond Distributed Transactions" [T04-S040]; Garcia-Molina+Salem 1987 Saga [T06-S023].
-
如果选型有 5 个候选: 不开会拍板; POC 2 个候选 2 周, ADR 记录 trade-off + 退出策略. Case: Ford+Richards 2020 Fundamentals of Software Architecture ch19 ADR 章 [T04-S023]; arc42 模板 [T06-S038].
-
如果监控只有 metrics: 加 traces (OpenTelemetry 标准) + structured logs + sampling; 三柱齐才能定位 cross-service 问题. Case: Google SRE book 4 Golden Signals [T03-S039]; Charity Majors observability 三柱 [T01-S038].
-
如果 LLM (Copilot / Cursor / Claude Code) 提议关键 trade-off: 默认拒绝, 自己想; LLM 仅用于 boilerplate / pattern lookup / ADR draft / test gen / commit message. Case: T03-D.2 LLM 红线; Veracode 2025 数据 45% AI 生成代码含 OWASP top-10 漏洞 [T03-S076]; Karpathy "vibe coding" 概念 [T03-S072].
-
如果 vendor 锁定 > 60% (业内估, 单一 cloud / 单一 db / 单一 observability): 标记 risk + 写 exit-strategy ADR + 至少 1 个 abstraction layer (OpenTelemetry / Terraform / SQL-only ORM) 隔离. Case: Gregor Hohpe Cloud Strategy (多 cloud / portable abstraction) [T01-S008]; T03-B.4 enterprise 反 vendor lock-in.
2.X Playbook 元规则 — 何时违背上述规则
- 永远不要把上述 10 条当圣经 — context-dependent 永远是 meta-rule. 银行 core 系统 (高合规 + 长 release cycle) playbook 完全不同于 创业 SaaS.
- 反模式: 把上面任何一条当 "best practice" 强推 — 反而创造新的 cargo cult. 用 playbook 是"先思考再决定 trade-off"的脚手架, 不是免思考 checklist.
工具栈与选型决策树
3.1 必备工具 (≥ 80% 当代 软件架构师用, 业内估)
- Diagramming / 系统设计 草图: draw.io (免费 + GitHub 集成) / Excalidraw (轻量) / Lucidchart (商业) / Miro / FigJam (协作白板). [T02-2.1]
- Architecture-as-code + ADR: Structurizr (Simon Brown C4 DSL) / PlantUML / Mermaid (Markdown 内嵌, GitHub 原生); adr-tools (Nygard) / log4brains. [T02-2.2]
- API design + spec: OpenAPI 3.x + Swagger Editor / Postman / Bruno (开源 alt); GraphQL: Apollo / Postgraphile; gRPC: protoc + buf.build. [T02-2.3]
- Database (OLTP 主流): PostgreSQL (default 选择) / MySQL / SQLite (单进程); Redis (cache + queue 兼用); MongoDB / DynamoDB (document); ClickHouse / DuckDB (OLAP). [T02-2.7]
- Observability: OpenTelemetry (vendor-neutral 标准) + Prometheus + Grafana (开源栈) 或 Datadog / Honeycomb / New Relic (商业 unified); Sentry (error tracking). [T02-2.8]
- Async messaging: Apache Kafka (大规模 streaming) / RabbitMQ (轻量 messaging) / NATS (cloud-native 轻量). [T02-2.4]
3.2 场景特化 (8 类, 按公司规模 / vertical 切)
Header 显式 "8 类" 让数字解析定位本节场景计数, 防 body 内 "T02" 等数字误抓.
- 创业 (5-15 eng, 1-3 services): 单体 + PostgreSQL + Redis + Fly.io / Render / Heroku; 不上 K8s / Kafka / Istio / service mesh. ADR 是 git markdown.
- Scale-up (50-200 eng, 10-30 services): modular monolith → bounded-context 拆 microservice; 引入 OpenTelemetry; service mesh 不急 (Linkerd 先于 Istio 复杂度低); ADR 进 repo.
- 大厂 (1000+ eng, 100+ services): K8s + service mesh (Istio / Linkerd) + Kafka + multi-region + 全 SRE + RFC + Backstage internal dev platform; principal engineer 网络.
- Enterprise / 银行 / 保险 / 电信 (重 governance): 重 compliance (SOX / GDPR / 等保 / 国密); vendor 优先 (Oracle / IBM / SAP); 中国: 信创 / 国密 / 国产化替代 (OceanBase / TDSQL / GaussDB / TDengine).
- 中国大厂高并发 (春晚级 / 双 11): 阿里 Dubbo + Sentinel (限流熔断) + Nacos (注册+配置) + RocketMQ + SkyWalking + OceanBase / TiDB; 蚂蚁 SOFAStack 在金融场景.
- Front-end + back-end 数据流: GraphQL Federation (大型多 client) / REST + OpenAPI (中小) / tRPC (TypeScript 单 repo monorepo); 状态层 React Query + Redux/Jotai/Zustand; BFF per-client 当 client 差异大.
- LLM-native 应用架构 (2024-2026 vertical): vector DB (pgvector / Pinecone / Qdrant) + LLM (Anthropic Claude / OpenAI GPT) + LangChain/LlamaIndex (谨慎用) + eval pipeline + LangSmith / Langfuse observability + cost-aware routing.
- AI 辅助架构师工作流: GitHub Copilot Chat / Cursor / Claude Code / Aider 提速 boilerplate 70-90% (业内估), 但 service boundary / polyglot persistence / trade-off articulation 不交 LLM.
3.3 新兴 / 实验 (近 12 月出现, 未必稳定)
- D5 Render (中国本土 AEC, 也开始用于架构 mood-board) — 软件架构无关, 跳过
- Cilium (eBPF-based service mesh, Istio 替代) [T02-2.5] — 2024 GA 但运维门槛高
- Cloud-native NewSQL TiDB Serverless / CockroachDB Serverless [T02-2.7] — 起步价低适合实验
- Backstage scaffolding plugin ecosystem (Spotify) [T03-S047]
- WebAssembly server-side runtime (wasmCloud / Spin) — 早期, 5-10% (业内估) 采用率
- AI eval frameworks (Braintrust / Helicone / LangSmith) — 6-12 月迭代速度
3.4 避坑清单 (外行 / 初阶 容易选错的)
- ❌ K8s for a CRUD app < 10 services — 运维复杂度 10x 产出 1x. 反例: 中型 SaaS 上 K8s + Istio + Prometheus + ArgoCD 全套, 6 月后回归 modular monolith on Render. Kelsey Hightower 自己说过 if you don't have a story for distributed systems failure modes, you should not be running Kubernetes [T01-Q-Hightower / T03-反模式].
- ❌ Service mesh first — Sam Newman 2nd ed do not start with service mesh. 反例: 10 服务规模上 Istio = 配置 maze + control plane overhead [T01-S014].
- ❌ Default microservices — 2014-2018 hype 余毒. Newman 自己 2nd ed 改口 [T01-S014]; Shopify 37 components modular monolith [T01-S058]; DHH 2025-10 X 长帖 "majestic monolith renaissance" [T01-S061].
- ❌ 多 service 共享 db — distributed monolith 经典标志. 每 bounded context 一个 db 才是 microservice 正经.
- ❌ 大厂方案抄到中小公司 — 阿里 Dubbo 全栈方案在 5 人公司 = 自杀. 99% (业内估) 中国软件公司应用 SpringBoot + MySQL + Redis 单体即可.
- ❌ ADR 写完不看 — waterfall ADR 反模式. ADR 必须 per-PR 嵌入 review 流程, 不是季度归档.
- ❌ LLM 决定 service boundary 或 polyglot persistence — 2024-2026 当前 hype 反模式. LLM 不知道你公司的团队 / 历史负债 / 长期演进, 跨系统 trade-off 必须人做.
工作流 / Pipeline
完整 SOP 见 [03-workflows.md] 三轴 (A 单系统 SOP / B 4 公司规模分流 / C 7 横向场景 / D AI 辅助). 本节摘要.
4.A 入门 SOP (创业期单系统设计, 1-3 月 cycle)
8 phase: 需求 → C1 Context → C2 Container → C3 Component → ADR+POC → 渐进上线 → 监控反馈 → 演进.
- W1: 需求理解 (1-2 周) — 利益相关者 + NFR (SLO / RPO / RTO / budget) + 业务用户故事 ≥ 5 条
- W2-3: C1 Context (1-2 周) — 系统 + 外部用户 / 系统 / 边界; 1 张 C4 Context 图; 5 个最大依赖
- W3-4: C2 Container (1-2 周) — 进程 / runtime / data store 边界; 1 张 C4 Container 图; 数据流向
- W5-6: C3 Component (2-4 周) — service 内组件 / 模块; 1 张 C4 Component 图 per service
- W6-8: ADR + POC (2-6 周) — 关键 trade-off (db 选型 / sync vs async / 一致性策略) 写 ADR; ≥ 2 选型 POC 2 周
- W8-10: 渐进上线 — feature flag + canary + dark launch; 5% → 25% → 100% traffic (canary 业内惯例)
- W10+: 监控反馈 — RED + USE + 4 Golden Signals; OpenTelemetry 三柱
- 持续: 演进 — fitness function 自动跑 + 季度 ADR 复审
- 资深路径: 跳过 W1-2 部分 stakeholder workshop (已有 NFR 模板) / 优化 W6-8 ADR (复用 arc42 模板, ≤ 1 页 lightweight) / 跳过 W5-6 单独 component 图 (合并 container 内 module 描述) / 额外补充 W10+ 的 fitness function 自动化 testharness, 不只 manual 走查.
4.B 资深路径 — 跳过 / 优化哪些步骤
- 已成熟 stack (e.g. JVM + PostgreSQL + Kafka) → 跳过部分 POC, 直接 ADR + 复用既有模板
- 已重组好的团队 (Team Topologies stream-aligned) → 跳过 RFC 跨组协调, 直接 squad 内决定
- 已有 fitness function library → 加新 service 时 inherit, 不重写
- 优化决策频率: 大决策季度 review (ADR refresh), 小决策 per-PR review, 不开会拍板
4.C 近期变化 (2024-2026)
- AI 辅助 boilerplate (Copilot / Cursor / Claude Code) 让 SD-implementation 缩短 30-50% (业内估), 但关键架构判断仍人做 [T03-D]
- LLM-native app architecture 是 2024-2026 新 vertical: RAG + agent loop + eval + cost-aware routing; 当前 hype 高点, 6-12 月会有 settling [T03-C.7]
- modular monolith renaissance — DHH "leaving the cloud" 2024+ / Shopify Packwerk / Stack Overflow case 在中文社区 2025-2026 也开始扩散 (反"默认微服务"风潮)
- 资深路径 (面对 2024-2026 变化): 跳过 "vibes coding" 全包 LLM 写代码 → 仅用 LLM 做 boilerplate / 优化 ADR draft 速度 (LLM 写初稿, 人改 trade-off 段) / 额外加 LLM-app eval pipeline + LangSmith/Langfuse observability 入 SOP
4.D 7 个横向场景 (cross-cutting workflow)
C.1 单体拆微服务 (strangler fig, 6-18 月) / C.2 全球多活 (multi-region active-active) / C.3 数据库选型决策树 / C.4 API gateway + BFF / C.5 Observability 落地 / C.6 Legacy refactoring (12-36 月) / C.7 LLM-native 应用架构. 完整见 [03-workflows.md C 节].
- 资深路径 (横向 case 共性): 跳过过早通用化 (don't build a platform until 3+ services need it) / 优化 strangler fig 节奏 (新功能必须只在新系统里 / 老系统不加新功能) / 额外用 fitness function 守护 "拆完不准回流" / 跳过自建 API gateway, 用 Kong / APISIX / Envoy 即可
表达 DNA
4 类 × 2-3 段 each = 10 段. 推断率 ≤ 30% (本 skill metadata 红线) (≥ 70% direct quote 或 转述自 documented source). 这些 voice samples 喂 Phase 4 voice check.
5.1 客户 / 业主 / 利益相关者对话 voice (需求 → trade-off 翻译)
5.1.1 (转述自 Hohpe Software Architect Elevator, T01-S008) — "业主问'什么时候上线?' 你不能答'看 backend 解耦完成度'. 翻译成业主语言: '如果接受双写期 2 月 + 旧 reader 兼容期 3 月, 8 月上线; 如果硬切, 4 月上线但回滚窗口只有 24h, 任何线上 bug 都是次日 page'." — 把技术语言翻译成"风险 + 时间表 + 退路"三件式.
5.1.2 (paraphrase, T03-A.1 需求理解 + T01-S008 Hohpe) — "我建议先 1 周做需求和 NFR brief, 不直接进选型. 因为 5 人团队和 200 服务大厂的'同样问题'答案完全不同. 你愿意花这 1 周吗?" — 在选型前强制 context 对齐, 不让业主以为架构是"选最好的工具".
5.2 同业 / 团队内对话 voice (ADR / RFC review)
5.2.1 (direct quote, Fowler MonolithFirst 2015 [T04-S007 / T01-S001]) — "Almost all the successful microservice stories have started with a monolith that got too big and was broken up. Almost all the cases where I've heard of a system that was built as a microservice system from scratch, it has ended up in serious trouble." — 拒绝 default microservices 时的标准引用.
5.2.2 (direct quote, DHH 2025-10 X 长帖 [T01-S061]) — "The majestic monolith renaissance is real. Cloud bills came due. Distributed system complexity came due. The pendulum is swinging back to 'just build it as one app and split when forced'." — 在 "为什么不拆服务" review 时直接 quote, 业内 anchor.
5.2.3 (转述自 Newman 2nd ed [T01-S014]) — "拆服务的 4 个合理理由: 不同 scale 需求, 不同 release 节奏, 不同团队 ownership, failure isolation 硬需求. 你这次想拆是哪个? 如果都不是, 这是 modular package 不是 service." — 拆 vs 不拆的 4-criterion 自查.
5.3 学术 / 教学 voice (讲解 CAP / Paxos / 一致性 / DDIA)
5.3.1 (direct quote, Kleppmann DDIA 2017 ch9 [T04-S035 / T01-Q-Kleppmann]) — "Consensus is the deepest problem in distributed systems. If you can solve consensus, you can solve almost any other problem in the field." — 解释为什么 Paxos / Raft 那么重要的标准开场.
5.3.2 (direct quote, Lamport Turing Award 2013 interview [T01-S026 / T01-Q-Lamport]) — "A distributed system is one in which the failure of a computer you didn't even know existed can render your own computer unusable." — 解释分布式 8 fallacies / 失败模型 的开场.
5.3.3 (paraphrase from Pat Helland "Life Beyond Distributed Transactions" [T04-S040]) — "在分布式里, 'transaction' 这个词只在一个 service 内成立. 跨 service 你能给的最强保证是 Saga + outbox + 幂等 receiver — 不是 2PC. 2PC 在生产环境就是个死刑." — 解释为什么不要做 2PC / XA 跨服务事务.
5.4 反例 / 批判 voice (反模式 / 不软化背书)
必含 ≥ 2 段反例 voice, 绝不软化 — 这是行业自省真实声音, 是 SKILL.md 不沦为 ThoughtWorks 推广材料的护身符.
5.4.1 (direct quote, Joel Spolsky "Don't Let Architecture Astronauts Scare You" 2001 [T01-S027 / T01-Q-Joel]) — "These are the people I call Architecture Astronauts. It's very hard to get them to write code or design programs, because they won't stop thinking about Architecture." — 反 over-abstract / 脱离 deliverable 架构师 的标准引用. 不要洗成"过分热心思考" — 原文就是嘲讽.
5.4.2 (direct quote, Joel Spolsky 同上 [T01-S027]) — "When you go too far up, abstraction-wise, you run out of oxygen. Sometimes smart thinkers just don't know when to stop, and they create these absurd, all-encompassing, high-level pictures of the universe that are all good and fine, but don't actually mean anything at all." — 反"高大上抽象设计图"的反例 voice.
5.4.3 (paraphrase from Sam Newman 2nd ed Building Microservices ch1 序言 [T01-S014] + Shopify modular monolith blog [T01-S058]) — "我们 2018-2022 把单体拆成 200 个微服务, 然后用了 18 月又合并回 37 个模块. 拆错的代价不是浪费 18 月, 是这期间业务功能 release 速度降到 1/3. 反过来如果一开始就 modular monolith + Packwerk 边界, 整个公司不会经历这场过山车." — 拆错的真实代价, 不软化为"经验教训".
5.4.4 (direct quote 转述, Kelsey Hightower [T02-2.6 / T01-Q-Hightower]) — "If you don't have a story for distributed systems failure modes, you should not be running Kubernetes. Monoliths are not the past — they are the future for most teams." — 反 K8s-for-everything, 出自 K8s co-creator 本人, anchor 极强.
质量基准 + 反模式
6.1 什么算"好"软件架构 (3-5 条可验证基准)
- 基准 1: 边界对齐 (boundary alignment) — service / module 边界跟"未来 12 月各自变化轴"对齐, 不是按当下功能流程切; 也跟团队 ownership 对齐 (Conway / Team Topologies). 验证: 看 commit log + 改一个 feature 涉及几个 service.
- 基准 2: trade-off 显式化 (explicit trade-offs) — 每个 key 决策有 ADR, 含 "what / why / alternatives / consequences / exit strategy" 5 段. 验证: 抽 5 个最近 6 月架构决定, 看 ADR 找得到吗.
- 基准 3: fitness function 自动跑 (continuous validation) — 至少 1 个自动化测试守护"什么算好" (e.g. ArchUnit 检查 dependency direction / 性能 budget / API breaking change detection). 验证: CI 跑 fitness 测试, 失败时 fail PR.
- 基准 4: observability 三柱齐 (metrics + traces + logs) — OpenTelemetry 标准 + RED + USE + Golden Signals; 不是只盯 Datadog dashboard. 验证: 一个跨 5 service 慢调用能 5 分钟定位到 root cause.
- 基准 5: context-aware (no cargo cult) — 不抄大厂方案到中小公司; 不把"best practice" 当圣经. 验证: 问 architect "你为什么这样选" 时, 答案是 trade-off 不是 "industry standard".
6.2 反模式 (反例 / 入门常犯的 7-10 错)
必须保留, 不软化 — 都是行业真实教训.
- Premature microservices — Default microservices 不问 4-criterion 直接拆 (T03 反模式 / Newman 2nd ed [T01-S014]).
- K8s for CRUD app < 10 services — 复杂度 10x, 价值 1x (Hightower 引用 [T01-Q-Hightower] / DHH "leaving the cloud" [T01-S061]).
- Service mesh first — 10 服务规模上 Istio = 配置 maze (Newman 2nd ed [T01-S014]).
- Over-engineering up-front design — 把所有未来 5 年场景设计进 day-1 架构; 真实业务 12 月内 70% (业内估) 假设会变. Architecture Astronaut 经典 (Joel Spolsky 2001 [T01-S027]).
- Waterfall ADR (写完落档) — ADR 写完归档没人读 = 0 价值. 改 lightweight + per-PR 嵌入 review.
- Fitness function ad-hoc — 只 manual 跑一次不进 CI; 等于没写 (Ford+Parsons 2017 [T04-S045]).
- Inverse-Conway 半截 — 重组 team 但没切服务边界; 或切服务但没重组 team — 都是 distributed monolith 加 reorg trauma 双煞.
- Vibes coding / LLM 决定关键 trade-off — 2024-2026 当前 hype 反模式 (Karpathy 创词 [T03-S072]); Veracode 2025 数据 45% AI 生成代码含 OWASP top-10 漏洞 [T03-S076].
- 大厂方案抄到中小公司 — 阿里 Dubbo 全栈在 5 人公司 = 自杀; ThoughtWorks Tech Radar 在 5 人公司也大多 mis-apply (Tech Radar 偏 ThoughtWorks 已咨询的中大型客户).
- 多 service 共享 db — distributed monolith 经典标志; 每 bounded context 一个 db 才是 service 边界正经 (DDD strategic + Newman [T04-S018 / T01-S014]).
智识谱系
每个流派: 奠基 → 当代代表 → 核心分歧.
| # | 流派 | 奠基 | 当代代表 | 核心分歧点 |
|---|
| 1 | Pattern movement | Fowler PoEAA 2002 / Hohpe EIP 2003 / GoF 1994 | Mark Richards / Alex Xu (ByteByteGo) | pattern catalog 是脚手架还是僵化模板? |
| 2 | DDD (Domain-Driven Design) | Evans Blue Book 2003 | Vernon (Red Book) / 张逸 (中文) | 大型团队 DDD 重 vs 启动期轻; tactical pattern 是否过度 |
| 3 | Clean / Hexagonal / Onion | Uncle Bob Clean Architecture / Cockburn 2005 hexagonal | Vlad Khononov / Mark Richards | TDD = professionalism 争议 (Uncle Bob 派 vs Coplien 2007 派) |
| 4 | Microservices | Newman 2014 / Richardson microservices.io | Newman 2021 2nd ed (自己改口) / Cockcroft | "default microservices" vs "modular monolith first" |
| 5 | Event-driven & log-centric | Young CQRS+ES 2010 / Kreps "The Log" 2013 / Helland | Vaughn Vernon (reactive integration) | EventSourcing 是否过重; cdc vs eventsourcing |
| 6 | Distributed systems thinking | Lamport Paxos 1998 / Kleppmann DDIA 2017 / Vogels | Charity Majors (observability) / Cindy Sridharan | formal methods (TLA+) vs practical engineering |
| 7 | Evolutionary architecture | Ford+Parsons 2017 Building Evolutionary Architectures | Neal Ford / Rebecca Parsons / Simon Brown (C4) | fitness function 工程化 vs ad-hoc |
| 8 | Cloud-native | Burns Designing Distributed Systems 2018 / Hightower | CNCF 项目集 (K8s / Envoy / Istio / OpenTelemetry) | K8s for everything (反模式 / Hightower 自己反) |
| 9 | Front-end architecture | Abramov Redux 2015 / Pete Hunt React 2013 | Dan Abramov (RSC) / Tanner Linsley (React Query) | micro-frontend Module Federation 是否过度 |
| 10 | 中国大厂中间件实践 | 陈皓 CoolShell 2009-2024 / 阿里中间件团队 | 李运华 / 沈剑 / 蔡超 / 张逸 (DDD 中文) | 大厂方案是否可移植到中小公司 (普遍 NO) |
Anti-figures (反例 / 不软化背书)
- "Architecture Astronaut" — Joel Spolsky 2001 创词 [T01-S027]. 反 over-abstract / 脱离 deliverable 的架构师.
- "Cargo cult microservices" / "Resume-driven development" — community 创词反模式. Stack Overflow 9-server case / Shopify modular monolith / DHH "leaving the cloud" 2024-2025 系列都是反"模仿 Netflix" 的真实反例 [T01-S058 / T01-S061 / T01-S062].
Sub-skills (女娲蒸馏的 top figures, Phase 3.3 选)
| Figure | 选理由 | sub-skill 路径 (建议) | 何时调用 |
|---|
| Martin Fowler | 软件架构 evergreen #1, 30 年 blog + 多本书 + Refactoring + PoEAA + 微服务系列 — long-form 最厚 + 跨流派思想 | sub-skills/martin-fowler/ | 当问题涉及 enterprise pattern / 重构 / 微服务取舍 |
| Martin Kleppmann | DDIA 2017 已成 2017-2026 必读 #1, 学术 + 实操桥梁, 长 talk + 论文 + open-source Automerge | sub-skills/martin-kleppmann/ | 当问题涉及分布式数据 / 一致性 / 数据流 |
| 陈皓 (左耳朵耗子) | 中文软件架构 #1 布道, 2023-05-13 已故但 CoolShell archive 长效 + MegaEase open source, 反 大厂崇拜 立场鲜明 | sub-skills/chen-hao/ (in memoriam 2023-05-13) | 当问题在中文语境 / 反大厂方案 mis-apply / 中文架构传播 |
注: 本 prototype run 不展开 sub-skill (nuwa-skill 未安装, 同前几轮). 后续如装 nuwa, 这 3 个候选直接用.
诚实边界
-
"架构师" 头衔模糊 — 同一 title 在不同公司含义完全不同 (Amazon principal engineer / 阿里 P8/P9 / 创业 founding architect / 咨询 architect / Gartner EA / AWS SA 售前). 本 skill 默认指 "落地 design + ADR + 跨服务边界 trade-off" 的从业者, 不含 Gartner 风格 EA / AWS SA 售前 / TOGAF / Zachman 框架训练.
-
Best practice 是 context-dependent — 创业期 (5-15 eng) ≠ scale-up (50-200 eng) ≠ 大厂 (1000+ eng) ≠ enterprise (银行 / 保险 / 政府); 同样问题 4 种公司答案不同; 不能笼统说 "always use X". 反例: 上面 10 条 playbook 都附 "context dependent" 元规则.
-
中文 canon 厚度差距 — 全球 canon 90% 英文奠基 (Parnas 1972 / Brooks 1975 / Fowler / Hohpe / Evans / Kleppmann); 中文当代 (李运华《从零开始学架构》/ 张逸 / 蔡超 / 沈剑) 主要 2015+ 起来, 厚度 中文:英文 ≈ 1:3-1:4 (业内估). 这是行业现实, 不是偏见.
-
ThoughtWorks Tech Radar / Gartner / Forrester 有 bias — ThoughtWorks 是 consulting firm, Tech Radar 倾向推自己擅长 / 已咨询客户用的技术 (例: 偏 microservices / DDD / Trunk-based dev); Gartner MQ 偏付费大厂; Forrester Wave 同样. 引用时标 "X 视角, 非全行业共识".
-
99% 中国软件公司 ≠ 阿里腾讯字节 — 阿里 Dubbo + Sentinel + Nacos + RocketMQ + SkyWalking 全栈是大厂规模本土实践; 中小公司是 SpringBoot + MySQL + Redis 单体 + 1 台 ECS. 中文社区文章 99% (业内估) 写大厂经验导致中小公司 mis-apply 是常见反模式.
-
AI / LLM 对架构师角色重塑是 2024-2026 hype 高点 — Copilot / Cursor / Claude Code 让 boilerplate / ADR draft / pattern lookup 接近免费; 架构师价值更集中在跨系统判断 + trade-off articulation + business ↔ tech 翻译. 但 "AI 架构师" 当前 hype 高点, 6-12 月 decay; 不在心智模型层做长期承诺.
-
微服务 / K8s / mesh first 反模式真实存在 — Newman 2nd ed 自己改口; Stack Overflow / Shopify / DHH 公开反例. 不要 default 这三个; 反过来 default modular monolith first.
-
TOGAF / Zachman 风格 enterprise architecture 不在本 skill 范围 — 那是 Gartner / 大企业 IT 治理 vertical, 与 application architecture 重叠 < 20% (业内估). 如果你 day-job 是企业架构师 / 解决方案架构师 (AWS SA 售前 / 咨询), 本 skill 仅参考价值.
-
Observability vendor 选型有商业拉锯 — Honeycomb vs Datadog vs New Relic vs Grafana stack 各有强弱; OpenTelemetry 是 vendor-neutral 标准但 backend 选什么仍有商业考量. 不背书单一 vendor; vendor lock-in 是 ADR 必谈话题 (playbook #10).
Time-decay Registry
This skill's modules decay at different speeds. Re-run update 大师 {slug}
when the dates below cross the recommended cadence (see references/extraction-framework.md § 八).
| Module | last_updated | decay_risk | Recommended refresh cadence |
|---|
| Mental models | last_updated: 2026-05-17 | decay_risk: low | 1-2 years |
| Standard playbook | last_updated: 2026-05-17 | decay_risk: low | 6-12 months |
| Tool stack | last_updated: 2026-05-17 | decay_risk: high | 3-6 months |
| Workflows / pipeline | last_updated: 2026-05-17 | decay_risk: high | 3-6 months |
| Expression DNA | last_updated: 2026-05-17 | decay_risk: low | 6-12 months |
| Sources (Track 5) | last_updated: 2026-05-17 | decay_risk: medium | 6 months |
| Glossary / standards / regulations | last_updated: 2026-05-17 | decay_risk: medium | 6 months (regulations may force sooner) |
| Intellectual genealogy | last_updated: 2026-05-17 | decay_risk: low | 1-2 years |
| Honest boundaries | last_updated: 2026-05-17 | decay_risk: low | re-assess each refresh |
last_updated values reflect the synthesis date. Individual research notes in
references/research/ may have more granular last_checked dates per item.
