Data Governance in the Era of AI, 2nd Edition (Early Release) (Evren Eryurek, Uri Gilad etc.)（Z-Library）

【One-Line Pitch】 A practical, lifecycle-spanning playbook for making enterprise data trustworthy enough to feed AI systems—and for governing what those systems do with it. Best for data leaders, governance and platform teams, and architects who must turn compliance obligations into working infrastructure. 【Book Arc】 - **Opening (~0%–9%)**: Frames why governance must enable rather than block AI, and introduces the shift that reshaped every governance tool: ML and generative AI as first-class consumers and producers of enterprise data, plus the new artifact of vector embeddings. - **Early (~9%–33%)**: Builds the vocabulary layer—enterprise dictionary, infotypes, data classes (PII, financial, business IP), and the enterprise policy book—and shows how policies attach to meaning rather than to individual tables or columns. - **Middle (~33%–52%)**: Moves into tooling: the dictionary/policy book/catalog triad, per-use-case policies, automated classification (pattern matching versus LLM-based), vector-store metadata and lineage, and open table formats as governance infrastructure. - **Late (beyond ~52%)**: Excerpts gesture at transformation and lineage questions—whether policies survive when data becomes a new representation—but the sampled material stops here; later chapters on quality, data in flight, protection, monitoring, and culture are listed as unavailable. 【Key Takeaways】 - **Governance is an enabler, not a brake** (Opening): the book's framing insists that trustworthy data is what makes AI adoption possible, not what slows it down. - **AI is both tool and object of governance** (Opening): the same LLMs that classify and catalog data are themselves systems whose training data, prompts, and outputs must be governed—and they widen the attack surface. - **Define data by meaning, then attach policy to meaning** (Early): policies tied to data classes ("junior analysts cannot access PII") scale and stay human-readable, unlike policies tied to a specific column. - **The dictionary, policy book, and catalog are three distinct artifacts** (Early): the dictionary defines words, the policy book writes the rules, and the catalog is where rules meet real tables, files, streams, and vector stores. - **Use case belongs in the policy** (Middle): role-based access alone is insufficient—the same address is fine for delivery but not for marketing, and multi-role employees break simple role models. - **Classification is going tiered and AI-assisted** (Middle): cheap pattern matching handles the obvious cases; LLMs catch nuanced meaning, but are slower, costlier, and less deterministic—and read every record they inspect. - **Embeddings need their own governance metadata** (Middle): record the producing model, chunking method, and a link back to source documents so lineage from governed record to vector is never lost. - **Open table formats are governance infrastructure** (Middle): Iceberg, Delta Lake, and Hudi bring schema evolution, time travel, and ACID transactions—time travel being the most consequential for point-in-time auditing and reproducibility. 【Reading Tips】 - Treat the early dictionary/class/policy material as the conceptual spine; if your organization lacks a dictionary, follow the book's advice and start with the highest-risk classes (PII first) rather than attempting exhaustive naming. - Deep-read the classification and vector-governance sections if you work with unstructured data or RAG pipelines—these are where the second edition diverges most from classic governance texts. - Skim the tool-name passages (DLP, Macie, catalog services) for capability categories rather than vendor specifics; the excerpts emphasize what to look for, not which product wins. - Watch for the recurring "AI as tool and object" thread—it is the book's organizing idea and reappears in nearly every chapter. 【Coverage Limits】 This guide is based on stratified excerpts covering roughly the first half of an Early Release edition; chapters on data quality, data in flight, protection, monitoring, and privacy culture are listed but unavailable, so later-stage content is not represented here.