The best metadata management examples come from teams that published their own numbers. Spotify (2020) lifted Lexikon adoption among data scientists “from 75% to 95%”. Atlan customer CME Group, which presented at Re:Govern 2025, reports a count of “over 18 million assets and 1,300+ glossary terms” cataloged in its first year. Shopify (2020) cut the share who said discovery hindered their work from 80% to 41%. Below are eight programs at six organizations, the layers each managed, and what kind of figure each published.
The in-house programs here wrote their own entity models, crawlers, search and lineage extraction between them: the core of metadata management and of a data catalog. Atlan ships several of those parts as a product. Glossary terms attach to any data asset, owners and tags propagate along lineage, and Atlan MCP serves that context to AI agents.
- Four layers recur (technical, business, operational, social), from a music service to a federal statute.
- Three programs published a before-and-after. Two published a relative change, and three a scale count.
| Field | Detail |
|---|---|
| Who this is for | Heads of data and analytics running a metadata program; AI leaders who need it ready for agents |
| Key regulations and standards | 44 U.S.C. 3511 federal data inventory requirement; DCAT-US metadata schema (v1.1 still harvested during the transition, 3.0 current) |
| Primary stakeholders | CDO or head of data platform, data governance lead, analytics engineering, privacy and legal |
| Typical data challenges | Data nobody can find, labels that differ by system, lineage drawn by hand, definitions kept apart from the assets |
| Data maturity level | From discovery programs (2020 write-ups) to classification and lineage at Meta’s scale (2025) |
| Fact | Detail |
|---|---|
| Programs covered | 8 programs at 6 organizations: Uber, Spotify, Shopify, Facebook/Meta (three programs), US federal agencies, CME Group |
| Years published | 2020 to 2025, plus a 2019 federal statute and a live federal catalog |
| Layers that recur | Technical, business, operational, social |
| Largest documented scale | Meta: “over 100 million schemas across more than 100 data systems” described in one format (count, 2025) |
| Most common success metric | Share of target users active weekly or monthly |
| Data.gov catalog size | 535,284 datasets as of November 2, 2026, including non-federal publishers |
What are the best examples of metadata management in large enterprises?
The best examples are programs whose owners published the problem, the metadata they managed, and a figure. Eight programs at six organizations meet that bar. Each names the company and the year; for catalog examples grouped by sector, see data catalog examples by industry.
| Organization (year published) | Problem in their words | Metadata they managed | Layers | How | Published figure (type) |
|---|---|---|---|---|---|
| Uber, Databook (2020) | “Databook was unable to scale to meet these new needs” | Datasets, dashboards, business metrics, pipelines, ML features, tables and their columns | T B O | Committee-approved schemas generate the data model; crawlers plus event-based sources | Before-and-after: new entity in “less than one hour” vs “multiple weeks” |
| Spotify, Lexikon (2020) | Discovery still “a major pain point” | Owners, descriptions, usage, lifecycle, joined tables, people and team pages | T B O S | Usage-weighted ranking, recommendations, expertise search | Before-and-after: adoption “from 75% to 95%”; MAU “~550 to ~870”; sessions per MAU “~3 to ~9” |
| Shopify, Artifact (2020) | Discovery ran through colleagues, chat, code and job logs | Documentation, schema, ownership, usage, lineage | T B O S | Generic data model; pull-based ingestion; lineage extractor | Before-and-after: “hindered” 80% to 41%; impact understanding 16% of the team to 46% of teams |
| Facebook, Nemo (2020) | Finding data “was a major pain point for data engineers” | Creator, columns, descriptions, lineage, usage, freshness, quality | T B O S | Ranking on trust signals; daily bulk plus instant indexing; learned re-ranking | Relative: search success up “more than 50 percent” as artifacts “more than tripled” |
| Meta, data understanding (2025) | Inconsistent labels across “hundreds of data systems” | Canonical schemas, semantic types, privacy taxonomy labels | T B | One schema format; classifiers with confidence thresholds; developer review | Count: “over 100 million schemas” described; annotation automated on “millions of assets” |
| Meta, lineage (2025) | Tracing data across “billions of lines of continuously evolving code” | Source and sink pairs, column-level lineage | T O | Static analysis plus sampled runtime probes | Relative: effort cut “by orders of magnitude” (no numeric baseline) |
| US federal agencies, Evidence Act inventories (2019) | Duty to “develop and maintain a comprehensive data inventory” | Descriptions, definitions, owner, access, last updated, restrictions | B O | DCAT-US catalog files harvested to Data.gov | Count: Data.gov lists 535,284 datasets (November 2, 2026), non-federal publishers included |
| CME Group, on Atlan (Re:Govern 2025) | Session summary: “business context took weeks to apply” | Assets, glossary terms, lineage | T B O | Atlan customer | Count: “over 18 million assets and 1,300+ glossary terms” in the first year |
Layer key: T technical, B business, O operational, S social.
Read the figure type before you borrow a number. A count shows reach. Only a before-and-after shows a program changed how people worked, so that is the one to copy into a metadata layer for AI plan.
Metadata management examples for data discovery and trust
Four programs attacked one problem: people who could not find or trust data. All four write-ups date from 2020. Treat them as snapshots, and check whether a tool still runs in that form before you copy its design.
Uber Databook (2020): an entity model that outgrew tables
Uber’s original Databook back end was built around datasets. According to Uber (2020), “Databook was unable to scale to meet these new needs”, because the old back end could not connect datasets, business metrics and dashboards.
The fix was a standard vocabulary for metadata: domain experts submit a schema definition, a central committee approves it, and Databook generates the data model from it. Crawlers and event-based sources feed it. Business metrics became entities next to the datasets and dashboards they relate to. That is the idea behind keeping KPI definitions and a business glossary in a metadata knowledge graph.
The before-and-after: Uber can now “onboard new entities in less than one hour”, where “it took multiple weeks” before. The platform covers “hundreds of thousands of datasets, millions of columns/fields.”
Spotify Lexikon (2020): usage signals turned search into recommendations
According to Spotify (2020), after Lexikon’s first release data scientists “still reported data discovery as a major pain point.”
Version two weighted search by popularity, recommended tables on a personalized homepage, and added expertise search built from what each person produces and consumes. 20% of monthly active users reached tables through personalized recommendations on the homepage. New entity pages, including people and team pages, drew “44% of Lexikon’s monthly active users.”
Lexikon adoption among data scientists rose “from 75% to 95%”. Monthly active users grew “from ~550 to ~870”, and sessions per user went “from ~3 to ~9”. Discovery was “no longer identified as a primary pain point.” Most of these changes drew on usage signals, which is worth remembering when you compare agents and humans in data discovery.
Shopify Artifact (2020): lineage so teams could see who their changes break
Before Artifact, Shopify’s data team found data by asking colleagues, searching chat, reading code and digging through job logs.
Shopify built a generic data model, pull-based ingestion, and a lineage extractor that separates direct from indirect dependencies. GraphQL APIs served the Artifact UI. Serving metadata to other teams through APIs was on Shopify’s mid- to long-term roadmap. That dependency map is the raw material for data lineage for AI.
According to Shopify (2020), “80% felt the pre-Artifact discovery process hindered their ability to deliver results. This sentiment dropped to 41%.” Shopify reports that 16% of the team felt they understood how their changes affected other teams, and that 46% of teams now do. The two figures use different units. Usage reached “over 30% of the Data team” weekly, “with a monthly retention rate of over 50%.”
Facebook Nemo (2020): trust signals when text search stopped scaling
At Facebook, finding the right data “was a major pain point for data engineers,” and plain text search got worse as the number of artifacts grew.
Nemo ranks results using trust signals drawn from data quality, recency, usage and lineage. It updates the index immediately when data changes, on top of a daily bulk load, and a learned model re-ranks results in a post-ranking phase.
According to Facebook (2020), Nemo increased “the data search success rate by more than 50 percent, even as the total number of artifacts has more than tripled”, and is “used by tens of thousands of engineers.” Of the four, Nemo is the one to study if your search already exists and keeps getting worse.
Metadata management examples for privacy and regulatory inventories
When a privacy duty or a statute drives the program, the job shifts from search to proving every asset is labelled, owned and traceable.
Meta (2025): one schema format and one taxonomy across 100+ systems
Meta faced inconsistent labels and missing annotations across “hundreds of data systems and millions of assets,” according to Meta (2025).
It defined one standard schema format, DataSchema, along with semantic types and a privacy taxonomy. Heuristic and machine-learning classifiers assign a label automatically when confidence clears a threshold, and developers verify the rest. Shared types like these are what handling PII in pipelines and data privacy for AI agents depend on. They are also a working form of ontology in AI.
DataSchema “describes over 100 million schemas across more than 100 data systems”, and annotation is automated “on millions of assets (i.e., fields, columns, tables).”
Meta (2025): lineage read from code at scale
Tracing sensitive data by hand stops working across “billions of lines of continuously evolving code,” according to Meta (2025).
Meta combined static code analysis with sampled runtime probes and static analysis of logged SQL to record column-level source and sink pairs. It reports “significantly reducing engineering efforts by orders of magnitude” but publishes no numeric baseline. Lineage captured from code is also what lineage over MCP and root-cause analysis through lineage need.
US federal agencies (Evidence Act, 2019): the inventory as a metadata spec
Under 44 U.S.C. 3511, as amended by the Evidence Act on January 14, 2019, each agency head “shall, to the maximum extent practicable, develop and maintain a comprehensive data inventory.” Agencies express it as DCAT-US catalog files, which Data.gov harvests. The Data.gov catalog listed 535,284 datasets on November 2, 2026, non-federal publishers included, so the total does not measure federal inventories.
The statute asks for “all variable names and definitions”, the owner, and the date last updated. That makes it a usable checklist for what metadata is supposed to carry. The descriptive fields searchers expect (title, description, keyword) land in the technical and business layers. The administrative ones (publisher, access level, rights, last modified) land in business and operational.
| Statutory item (44 U.S.C. 3511) | DCAT-US v1.1 field | DCAT-US 3.0 note | Layer |
|---|---|---|---|
| Description, including all variable names and definitions | description; variable definitions through describedBy, the data dictionary URL |
description still present in 3.0 |
B |
| Name or title | title |
Still present in 3.0 | T |
| Whether the asset is open, disclosable or public | accessLevel (for example "public") |
Not part of the 3.0 core schema; 3.0 uses accessRights |
B |
| How the public can access or request access | distribution with accessURL / downloadURL |
Not covered by the 3.0 notes cited here | O |
| Date most recently updated | modified |
Present; listed among fields most likely to fail 3.0 validation | O |
| Each agency responsible for maintaining it | publisher; bureauCode for federal agencies |
publisher still present; bureauCode accepted but outside the 3.0 core schema |
B |
| Owner | contactPoint |
Still present in 3.0 | S |
| Restrictions on use | rights |
Free-text restrictions move to AccessRestriction, UseRestriction and CUIRestriction classes | B |
| Location | accessURL / downloadURL |
Not covered by the 3.0 notes cited here | T |
The mapping is our reading of the statute against the schema, not an official crosswalk. Field names are DCAT-US v1.1. DCAT-US 3.0 is the current version, and “v1.1 continues to be harvested by Data.gov during the transition period,” so v1.1 catalog files like the one below are still live.
{
"conformsTo": "https://project-open-data.cio.gov/v1.1/schema",
"dataset": [{
"title": "Example Dataset",
"description": "Placeholder description including variables and their definitions.",
"keyword": ["example"],
"modified": "2026-10-01",
"publisher": {"name": "Example Agency"},
"contactPoint": {"fn": "Data Owner", "hasEmail": "mailto:[email protected]"},
"identifier": "example-agency-0001",
"accessLevel": "public",
"bureauCode": ["010:86"],
"programCode": ["015:001"]
}]
}
Both programs need a shared definition on every asset; the statute adds an owner and a last-updated date.
How did these programs measure metadata management success?
The programs with a published before-and-after measured against a baseline taken before the new version shipped. Spotify and Shopify each published an adoption figure and an outcome figure. Shopify went one step further and published its targets: it “exceeded our expectations of 20% of the Data team using the tool weekly, with a 33% monthly retention rate.”
| What you want to prove | Metric the program used | Who used it | Before and after published? |
|---|---|---|---|
| People find data | Share of target users active weekly or monthly | Spotify 2020 (75% to 95% of data scientists), Shopify 2020 (over 30% weekly against a 20% target) | Yes |
| People come back | Retention or sessions per user | Spotify 2020 (~3 to ~9 sessions per MAU), Shopify 2020 (monthly retention over 50% against a 33% target) | Yes |
| Search actually works | Search success rate | Facebook Nemo 2020 (up more than 50%) | Relative only |
| Pain goes away | Survey sentiment on a fixed question | Shopify 2020 (80% to 41%), Spotify 2020 (no longer a primary pain point) | Yes |
| Change is safe | Share who understand downstream impact | Shopify 2020 (16% of the team to 46% of teams) | Yes, units differ |
| The model stays extensible | Time to onboard a new entity type | Uber 2020 (multiple weeks to under one hour) | Yes |
| Labelling keeps up | Assets classified automatically | Meta 2025 (annotation automated on millions of assets) | Count only |
| Coverage | Assets or schemas under management | Meta 2025 (100M+ schemas described), CME Group, Re:Govern 2025 (18M+ assets, 1,300+ terms), Data.gov (535,284 datasets, November 2, 2026) | Count only |
Sources for each row: Uber, Spotify, Shopify, Facebook, Meta, CME Group, Data.gov.
Copy Shopify’s habit: write the target and the baseline down before launch. Every metric above asks whether people found and trusted data. None asks whether a machine could answer a question from the same context. For AI readiness, add that one: give an agent a set of questions with known answers and record how many it gets right.
What goes wrong in metadata management programs, and how did teams fix it?
Six failure modes recur, each with the fix the team shipped.
| Failure mode | Who hit it (year) | Fix they shipped |
|---|---|---|
| The model only covers tables, so metrics and dashboards stay dark | Uber (2020) | Committee-approved schemas that generate the data model for new entity types |
| Search alone leaves discovery painful | Spotify (2020) | Usage-weighted ranking, recommendations, people and team pages |
| Text search degrades as assets grow | Facebook (2020) | Ranking on trust signals from quality, recency, usage and lineage |
| Nobody knows whose work a change breaks | Shopify (2020) | Lineage graph with direct and indirect dependencies |
| Labels differ across systems and manual annotation cannot keep up | Meta (2025) | One canonical schema, a shared taxonomy, classifiers with confidence thresholds plus developer review |
| Lineage drawn by hand goes stale | Meta (2025) | Static analysis plus sampled runtime probes |
Five of the six fixes derive context from the systems themselves (usage logs, code, pipelines, classifiers) rather than relying on people to type it all in. Uber’s is the exception: a central committee of experts approves each proposed schema. That split is where semantic understanding versus metadata management becomes a practical distinction.
Atlan follows the same pattern for the derived part. It propagates “owners, certificates, tags, and custom metadata” from upstream source assets along lineage, and its playbooks “automate metadata management and governance as repeatable, rule-based workflows.”
Building a metadata platform in-house vs. a shared context layer
Six of the eight examples are in-house builds: the four 2020 discovery platforms and Meta’s two 2025 programs. Between them, these teams wrote entity models, crawlers, search ranking and lineage extraction.
| Question | What the in-house programs did | What a shared context layer changes |
|---|---|---|
| Who defines entity types? | Uber’s committee approved schemas that generated the data model; Shopify wrote a generic data model | Teams define glossary terms and attach them to assets instead of writing the model |
| How does business meaning attach to assets? | Separate builds: Uber’s vocabulary, Meta’s taxonomy and semantic types | Glossary terms attach to any data asset |
| How is lineage captured and used? | Shopify’s extractor; Meta’s static analysis and runtime probes | Owners, tags and certificates propagate along lineage |
| Who consumes the context? | People in a UI (Shopify’s GraphQL APIs served its UI); Uber’s API also “supports other data tooling at the company” | People and AI agents; agents read governed definitions and lineage over MCP at query time |
The deciding question is staffing. Each of these platforms was built and run by the company’s own engineers. Teams weighing build versus buy for metadata tooling can set the cost of building it in-house against what changes from catalog to context layer and from active metadata to a context layer. Enterprise metadata management covers the operating side.
How does Atlan support metadata management programs like these?
Atlan provides several of the parts these programs built in-house, in one Enterprise Data Graph that serves people and AI agents.
CME Group shows the scale. Atlan’s Re:Govern 2025 session summary says that at CME Group “business context took weeks to apply.” On Atlan, CME Group “cataloged over 18 million assets and 1,300+ glossary terms in our first year”, according to the Context Lakehouse page. That figure is a count, with no before-and-after attached.
What Atlan provides against these programs’ parts:
- Definitions on the asset: in Atlan, “glossary terms can be attached to any data asset” (Atlan docs), the job Uber’s vocabulary and Meta’s taxonomy did.
- Lineage-driven metadata: owners, certificates, tags and custom metadata propagate through lineage from upstream assets.
- Curation by rule: playbooks.
- Agent access: the Atlan MCP server lets AI clients “search and discover enterprise context, traverse end-to-end lineage” and “access governed data definitions and glossaries” (Atlan docs). That read path separates an MCP-connected data catalog or an agentic data catalog from a search box.
Together these form an enterprise context layer, the base that context engineering and a semantic layer both draw on.
Which metadata management example fits your situation?
Match your driver to the closest example, build that example’s first layer, and baseline its first metric before you launch.
| Your driver | Closest example | First layer to build | First metric to baseline |
|---|---|---|---|
| Analysts cannot find or trust data | Spotify, Shopify, Facebook Nemo (2020) | Social (usage, owners), then business | Weekly active share; survey sentiment |
| The data model keeps outgrowing tables | Uber (2020) | Technical (an extensible entity model) | Time to onboard a new entity type |
| A privacy or erasure obligation | Meta (2025) | Business (shared definitions and a taxonomy for personal data) | Share of assets classified, with review |
| A statutory or regulator inventory | US federal agencies (2019) | Business plus operational (owner, last updated, access) | Inventory completeness against the required fields |
| Changes break downstream users | Shopify (2020) | Operational (lineage) | Share who understand downstream impact |
| Agents need governed answers | No published example yet | Business (definitions linked to assets) | Answer correctness on a fixed set of known questions |
To get started:
- Name the driver in one sentence, with the person who feels it.
- List the entity types in scope: tables only, or also dashboards, metrics, pipelines and ML features.
- Check whether a definition, an owner and a last-updated date live on each asset. Those are the types of metadata agents need most.
- Pick two metrics from the measurement table, set a target for each, and record today’s numbers.
Two pitfalls these programs document:
- Modelling tables only. Uber’s original back end (2020) could not hold business metrics and dashboards.
- Ranking search on text alone. Facebook’s search (2020) degraded until Nemo added trust signals.
If agents are the driver, read metadata management for AI and the role of metadata management in enterprise AI before you implement an enterprise context layer.
Real stories from real customers: Metadata at enterprise scale
"Context is the differentiator. Atlan gave our teams the shared vocabulary and lineage to move from reactive data management to proactive AI enablement."
Kiran Panja, Managing Director, Cloud and Data Engineering, CME Group
"AI initiatives require more context than ever. Atlan's metadata lakehouse is configurable, intuitive, and able to scale to hundreds of millions of assets. As we're doing this, we're making life easier for data scientists and speeding up innovation."
Andrew Reiskind, Chief Data Officer, Mastercard
What these eight programs leave you to measure
The strongest evidence here is old and specific. Spotify, Shopify and Uber published a baseline, so their gains can be checked. Meta, the federal inventories and CME Group show how far the same layers stretch, in counts.
The discovery programs measured people, such as who searched and who came back. None of the write-ups reports whether a machine could answer a question from the context. If your next consumer is an agent, that is the baseline none of these programs published, and the first one to record.
FAQs about metadata management examples
1. What is an example of metadata management?
Spotify’s Lexikon is one of the best documented. In 2020 Spotify reported that after it ranked datasets by usage, recommended tables and added people and team pages, Lexikon adoption among its data scientists rose from 75% to 95%.
2. What are the four types of metadata, with examples?
Technical metadata describes structure, like Meta’s single schema format (2025). Business metadata carries meaning, like Uber’s committee-approved entity schemas (2020). Operational metadata records how data moves, like Shopify’s lineage graph (2020). Social metadata captures people and usage, like Spotify’s team pages (2020).
3. What is an example of business metadata?
A definition attached to the asset it describes. Under 44 U.S.C. 3511, as amended by the Evidence Act in 2019, each agency’s data inventory must carry, to the maximum extent practicable, a description of every data asset including all variable names and definitions, plus its owner and the date it was last updated.
4. How do you measure whether metadata management is working?
Write down a target and a baseline before launch. Shopify (2020) set targets of 20% weekly use and 33% monthly retention, then reported over 30% weekly use and over 50% monthly retention. Uber (2020) tracked the time to add a new entity type, which fell from multiple weeks to under an hour.
5. Do you need to build an in-house metadata platform like Uber or Spotify did?
Only if you can staff a platform team for it. Uber, Spotify, Shopify and Facebook each built their own platform in-house, and each described it in 2020. Between them they wrote entity models, crawlers, search ranking and lineage extraction. Without that team, building is the expensive path, and a shared context layer moves the work to definitions and ownership.
6. How does metadata management support AI agents?
An agent needs the definition, owner, freshness and lineage on each asset, and a way to read them at query time. Atlan’s MCP server gives agents that read path. Through it they can search enterprise context, traverse lineage and access governed definitions.
Sources
- Turning Metadata Into Insights with Databook, Uber Engineering, November 10, 2020
- How We Improved Data Discovery for Data Scientists at Spotify, Spotify Engineering, February 27, 2020
- How We’re Solving Data Discovery Challenges at Shopify, Shopify Engineering, July 21, 2020
- Nemo: Data discovery at Facebook, Engineering at Meta, October 9, 2020
- How Meta understands data at scale, Engineering at Meta, April 28, 2025
- How Meta discovers data flows via lineage at scale, Engineering at Meta, January 22, 2025
- 44 U.S. Code § 3511, Data inventory and Federal data catalogue, Legal Information Institute, Cornell Law School
- DCAT-US Schema v1.1 (Project Open Data Metadata Schema), resources.data.gov
- DCAT-US Schema v3.0, resources.data.gov (accessed 2026-11-02)
- Catalog, Data.gov, US General Services Administration (count read 2026-11-02)
- A Business Graph, Atlan documentation (accessed 2026-11-02)
- Propagate metadata through lineage, Atlan documentation (accessed 2026-11-02)
- Playbooks, Atlan documentation (accessed 2026-11-02)
- Atlan MCP, Atlan documentation (accessed 2026-11-02)
- Context Lakehouse: The World’s First Context Store Built for AI, Atlan (first-party, accessed 2026-11-02)
- CME Group: Context at Speed, Re:Govern 2025, Atlan (first-party)
