In 2019, a global reinsurer discovered it had 14 different definitions of "net written premium" living across its warehouses, spreadsheets, and BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.Voir la définition complète → tools. The number that hit the board deck depended on which analyst pulled it, and which of the 14 they happened to trust that week. The company had spent nine figures on data infrastructure. It had a governance council, a data dictionary, and a well-staffed BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.Voir la définition complète → team. What it did not have was a way to answer a deceptively simple question: *where did this number come from, and can I trust it?*
That gap—between owning data and being able to find, understand, and trust it—is the difference between a data asset and a data swamp. And it is exactly the gap a modern data catalogdata catalogA centralized inventory of an organization's data assets, enriched with metadata, that helps people find, understand, and trust the data they need.Voir la définition complète →, powered by active metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition)., is built to close. Not through a prettier data dictionary. Through a fundamentally different operating model for metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition)..
You already know metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). as "data about data." The distinction that matters at your level is between passive and active metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition).—and it is the single most important framing you should carry out of this lesson.
Passive metadata is inventory. It's the schemaschemaA schema is the formal blueprint that defines how data is structured, named, typed, and related within a database, file, or message.Voir la définition complète →, the column names, the owner field someone filled in during a migration two years ago. It sits in a repository. It gets stale the moment it's written. The old-world data dictionary and the first generation of catalogs were passive: humans documented, humans searched, and the whole thing decayed at the speed of organizational neglect.
Active metadata is metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). in motion. It is continuously harvested from the systems that actually run your data estate—query logs, orchestration tools, BIBITechnologies and processes that turn raw data into actionable insights via reporting, dashboards and analysis, so teams can decide based on facts rather than intuition.Voir la définition complète → platforms, transformation frameworks, cloud APIs—and then *pushed back out* to the places where people work. The catalog stops being a library you visit and becomes an intelligence layer that acts on what it knows.
Gartner formalized this shift when it retired its "MetadataMetadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). Management Solutions" Magic Quadrant in 2021 and replaced it with the concept of active metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). management. The point wasn't a naming change. It was recognition that metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition).'s value is proportional to how *current* and how *actioned* it is.
Here is the practical test to apply on Monday morning. For any metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). your catalog holds, ask:
A catalog that only ingests is a museum. A catalog that ingests *and* activates is infrastructure.
Not all metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). is equally valuable. Prioritize the harvesting of four signal types, in roughly this order:
1. Technical metadata — schemas, types, partitions. Table stakes, automatically scanned.
2. Operational metadata — freshness, volume, query frequency, failure rates. This is where trust signals are born.
3. Lineage metadata — the dependency graph (covered in depth in the next lesson, but it lives in the catalog).
4. Business metadata — definitions, ownership, sensitivity classification, certification status.
The failure mode of most enterprises is inverting this order: pouring effort into a business glossary (item 4) while ignoring operational signals (item 2). You end up with beautifully documented tables nobody can tell are broken.
Vendors will pitch you a feature matrix. Ignore it. Judge a catalog against the *outcomes* it must produce for your organization. There are four, and you should hold any investment—build or buy—accountable to all four.
The core failure a catalog fixes is search. When an analyst can't find the right dataset, one of three bad things happens: they use the wrong one, they rebuild one that already exists (the shadow-copy epidemic), or they file a ticket and wait a week. Airbnb built its internal Dataportal precisely because analysts were spending more time hunting for data than analyzing it, and no one could tell which of several similar tables was authoritative.
Effective findability is not full-text search over table names. It is ranked, context-aware search: the most-queried, most-recently-updated, most-connected assets surface first, weighted by the searcher's team and role. This is where active metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). earns its keep—popularity and freshness are observed signals, not curated ones.
Finding a table called fct_revenue_daily is worthless if you don't know what a "day" means (booking date? recognition date? UTC?). Understandability means every asset carries its definition, its owner, its lineage, sample values, and known caveats—and that this context is one click from wherever the analyst is working.
The judgment call here is curation depth versus coverage. You cannot hand-document 200,000 tables. So you tier them. Certify the 3–5% that feed executive reporting and customer-facing products to a high standard. Let the long tail be auto-documented and community-annotated. A CDO who insists on uniform documentation quality across the whole estate is guaranteeing the project never finishes.
This is the reinsurer's problem. Trust is delivered through two mechanisms working together:
The magic is combining them. Certification without live signals is a stale promise. Signals without certification is noise no one knows how to act on. Together, they let an analyst answer *"can I trust this number?"* in seconds—and let you kill the 14-definitions problem at its root, because the certified asset visibly outranks the shadow copies.
Your governance module taught you the policy framework. The catalog is where policy stops being a PDF and becomes enforced reality. Active metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). makes classification *executable*: when the catalog tags a column as PII through automated scanning, that tag can propagate downstream through lineage and trigger masking, access restrictions, or retention rules automatically.
This is the connective tissue between the catalog and your existing governance and compliance investments. A tag that only informs is passive. A tag that *acts*—that revokes access or masks a field—is the payoff.
The tech giants built their own—Airbnb's Dataportal, LinkedIn's DataHub, Lyft's Amundsen, Uber's Databook—because their scale and engineering depth justified it, and several open-sourced the result. For nearly every other enterprise, building is a strategic error. You will spend two years reinventing connectors instead of driving adoption.
The real decision is between commercial platforms (Alation, Collibra, Atlan, Informatica) and mature open-source (DataHub, OpenMetadata). Decide on three axes:
Do not attempt a big-bang catalog of the entire estate. That is how these programs die. Sequence it:
1. Start with a high-pain, high-visibility domain. Finance reporting or customer analytics—somewhere the "which number do I trust" pain is acute and a win is visible to executives.
2. Auto-harvest everything technical and operational first. Get the machine-generated signals flowing before you ask a single human to document anything. This proves value with zero manual burden.
3. Certify the critical few. Put owners' names on the assets that matter most in your beachhead domain.
4. Instrument adoption from day one. Track catalog search sessions, click-through to assets, and—most importantly—reduction in "where's the data?" tickets and duplicate-dataset creation.
5. Expand domain by domain, using each success to fund the next.
The metric that predicts success is not tables cataloged. It is weekly active users among analysts. A catalog no one opens is a stranded asset, no matter how complete.
Vérification des acquis
1. The reinsurer's inability to answer 'where did this number come from, and can I trust it?' most fundamentally illustrates which distinction?
2. Why does the lesson argue that a modern data catalog is NOT simply 'a prettier data dictionary'?
3. What is the essential behavioral difference between passive and active metadata?
4. Select ALL correct answers about the characteristics of passive metadata as described in the lesson.
Sélectionnez toutes les réponses correctes.
5. Select ALL correct answers about the sources and behavior of active metadata.
Sélectionnez toutes les réponses correctes.
The most common way a catalog investment fails is treating it as a one-time implementation. You buy the tool, you run the scan, you declare victory, and eighteen months later the metadatametadataDonnées sur les données, informations décrivant le contexte, la structure, la provenance et les caractéristiques d'un asset de données (auteur, date, format, source, définition). has decayed and adoption has cratered.
A catalog is a living product with a permanent owner, a roadmap, and a feedback loop. Someone—usually a data governancedata governanceData governance is the set of policies, roles, and processes that ensure data is accurate, secure, well-defined, and used responsibly across an organization.Voir la définition complète → or platform lead reporting to you—owns catalog adoption as an ongoing KPIKPIKey Performance Indicator, a measurable value that shows how effectively you're achieving a specific objective, tracked over time against a target.Voir la définition complète →. They watch what people search for and fail to find, and they close those gaps. They enforce that new data products can't reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.Voir la définition complète → production without catalog registration and an assigned owner (wire this into your deployment pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète → as a gate, so it's automatic, not aspirational).
The stewardship model matters as much as the tool. Federated ownership—domain teams certify their own assets against central standards—scales; centralized ownership does not. This is the operational expression of the data-as-a-productdata-as-a-productA data asset managed like a product, with an owner, defined users, guaranteed quality, and measurable business value.Voir la définition complète → thinking you'll deepen elsewhere in this track: every dataset has an owner accountable for its catalog presence, just as a product manager owns a product.
The organizations that win treat catalog quality metrics—percentage of critical assets certified, percentage with assigned owners, freshness of documentation—as first-class governance KPIs reported alongside compliance and quality. Those that lose treat the catalog as IT plumbing and wonder why the 14 definitions came back.