云原生数据中台：架构、方法论与实践 2021 (彭锋 宋文欣 孙浩峰)（Z-Library）

【One-Line Pitch】 A practical field guide to building a cloud-native data middle platform (数据中台) that turns scattered big-data tooling into shared, reusable data capabilities. Best for data platform architects, big-data team leads, and technical managers who must decide how to evolve an existing warehouse/lake stack into something the whole company can actually use. 【Book Arc】 - **Opening (~0%–15%)**: Frames why "data middle platform" emerged in China while Silicon Valley firms built equivalent capability without the label, drawing on Twitter's platform history; defines the concept and its four characteristics, plus evaluation criteria. - **Early (~15%–33%)**: Covers the digital-transformation stages (informatization → data-driven), the three-step path from big-data platform to middle platform, common pitfalls, and the case for top-level architecture design with clear responsibilities. - **Middle (~33%–52%)**: Presents the architecture and methodology — OneID/OneModel/OneService/TotalPlatform/TotalInsight norms, cloud-native PaaS as the necessary base, observability, and pragmatic open-source selection. - **Late (~52%+)**: Moves into component-level practice: data lakes and ODS, interactive analysis and BI tooling (Zeppelin, Jupyter, Superset), and the build-out of data development and application layers. - **Ending**: The excerpts do not cover the closing chapters in detail, so the final synthesis and forward-looking material cannot be summarized here. 【Key Takeaways】 - **A data middle platform is defined by shared, abstracted, reusable data capability** (Opening): not a new tool but an operating layer above warehouses and big-data platforms, judged by standard coverage, reuse rate, and measurable ROI. - **Silicon Valley solved the same problem without the label** (Opening): Twitter's unified platform let product, ads, and anti-fraud teams reuse each other's data capabilities, keeping iteration cycles shrinking from months to weeks. - **Data-driven systems need three properties: continuous, insight-generating, dynamic** (Early): T+1-to-T+0 processing, analytics/ML rather than raw display, and outputs generated per-user rather than by fixed rules. - **Build in three stages, not one leap** (Early): platform construction, data management/application, then capability platformization — knowing your stage determines which problems to expect. - **Top-level design matters, but don't over-invest in upfront modeling** (Early): define master data, data domains, and ownership, then iterate from real business pain points rather than exhaustive surveys. - **Cloud-native PaaS is the necessary foundation** (Middle): microservices, containers, DevOps, and CI/CD let data applications inherit multi-tenancy, storage, load balancing, and fault tolerance instead of rebuilding them. - **Choose mature open-source cores, customize only the interaction layer** (Middle): prefer community-verified infrastructure; patch source only for early-stage projects or production-blocking bugs. - **Observability spans logs, metrics, and traces** (Middle): ELK-style logging, Prometheus-style metrics, and Zipkin/SkyWalking tracing — with service mesh providing call tracing for free. 【Reading Tips】 - Read Part 1 (Opening) closely for the definition and evaluation criteria — it is the lens for judging every later architecture decision. - Skim the Silicon Valley platform case studies (Twitter, Airbnb, Uber) for shared patterns rather than memorizing component names. - Treat the architecture and methodology chapters as the book's spine; the component chapters are reference material to revisit during actual selection. - Pay attention to the organizational and responsibility discussions — the book repeatedly stresses that data middle platforms fail for non-technical reasons. - Keep the open-source selection principles handy as a checklist when evaluating specific tools. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of the book; later component chapters and the ending are only partially represented, so tool-specific and concluding material is summarized only where the excerpts support it.