Quick answer
A data catalog is an organized inventory of an organization's data assets, built from metadata, that answers four questions about each asset: what data exists, where it lives, who owns it, and whether it can be trusted for a given use.
What is a data catalog?
A data catalog is an organized inventory of an organization's data assets, built from metadata. IBM defines it as "a detailed inventory of data assets within an organization" that helps users discover, understand and access data; AWS calls it "an inventory of all the data that an organization collects and processes". The practical test is whether the catalog answers four questions about any asset:
- What data exists? A name and a plain-language definition, not just a table name.
- Where does it live? System, database or project, table or event name, and how to query it.
- Who owns it? A named person or team responsible for the definition and the pipeline behind it.
- Can I trust it? Freshness, quality checks, known caveats, and its status (active, deprecated, still in development).
A catalog that answers only questions 1 and 2 is a list. Questions 3 and 4 are what turn the list into something analysts actually use — and, as the example below shows, what keeps it from going stale within a quarter.
The catalog is assembled from metadata, usually sorted into three types (AWS groups them this way): technical metadata (schemas, columns, formats, locations), business metadata (definitions, owners, classifications, business context), and operational metadata (update jobs, freshness, quality results, usage). Modern catalogs harvest technical metadata automatically — the AWS Glue Data Catalog, for example, is a persistent metadata store that crawlers fill with table definitions — while business metadata such as ownership almost always needs a human to enter and maintain it.
Data catalog vs. data dictionary vs. metadata management
These terms overlap, and vendors use them loosely. The table separates them by the question each one answers:
| Concept | Question it answers | Scope | Typical contents |
|---|---|---|---|
| Data dictionary | What does each field in this dataset mean? | One dataset or database | Field names, types, allowed values, units, meanings |
| Data catalog | What data exists, where, who owns it, can I trust it? | All assets across the organization | Inventory + dictionary layer + owners, freshness, lineage, usage |
| Metadata management | How is metadata collected, stored and kept current? | The machinery under the catalog | Harvesting, storage, synchronization, APIs |
| Business glossary | What does each business term mean? | Business language, not storage | Term definitions, synonyms, related metrics |
| Data governance | Who may decide and do what with data? | Policies and roles | Ownership rules, access policies, quality standards |
Two relationships are worth memorizing. The dictionary is a layer inside the catalog: IBM puts it plainly — the catalog gives a broad overview of all assets, while the dictionary "defines the structure and content of individual datasets". And the catalog is a tool of governance, not governance itself: governance decides that every asset must have an owner; the catalog is where that owner is recorded and found.
What a catalog entry contains
Here is one complete entry, using an analytics event rather than a warehouse table (sample data):
| Field | Value for this entry |
|---|---|
| Asset name | signup_completed (event) |
| Definition | Fires once when a user finishes account creation, after email verification succeeds. Excludes social-login auto-accounts. |
| Where | Product analytics project "Main app"; platforms web + iOS; query it in event analysis |
| Owner | Growth team — named analyst responsible for the definition |
| Properties | method (string, required: email / sso), plan (string, optional) |
| Status | Active; verified 2026-11-01 |
| Freshness / quality | Arrives in real time; required-property completeness checked daily |
| Caveats | Before 2026-08, iOS fired it before verification; historical counts are ~4% higher |
| Related assets | Feeds the activation funnel; upstream of the "new accounts" metric |
Notice what each field does: the definition prevents two teams from counting different things; the owner gives you someone to ask; status and the verified date tell you whether the entry itself is current; the caveat saves an analyst from a wrong historical comparison. Lineage — the "related assets" row — is its own discipline; see data lineage. For how the property rows should be structured, see event schema.
A worked example: three teams, three definitions of signup
Sample data. A company has 50,000 visitors in October. Three teams each built their own "signup" count, and none of them wrote down the definition:
- Marketing counts form submissions: 1,050.
- Product counts verified accounts: 1,050 − 170 unverified = 880.
- Finance counts accounts on paid plans: 880 − 640 free = 240.
The marketing and product dashboards both show a "visitor → signup" funnel:
| Team | Numerator | Funnel rate |
|---|---|---|
| Marketing | 1,050 form submissions | 1,050 ÷ 50,000 = 2.10% |
| Product | 880 verified accounts | 880 ÷ 50,000 = 1.76% |
Same month, same visitors, a 0.34-percentage-point gap — a 19% relative difference (0.34 ÷ 1.76) that will be blamed on "tracking bugs" in every meeting. Nothing is broken; the word "signup" just means three different things.
The fix is one catalog entry, not a new pipeline: signup_completed = verified account creation, owner Growth team, status active. Marketing renames its metric to signup_form_submitted (a different, legitimate entry), and both funnels that claim to measure signups now read 1,050 ÷ 50,000 = 2.10% for form submissions or 880 ÷ 50,000 = 1.76% for verified accounts — labeled, consistent, and no longer in conflict.
The second cost is search time. Before the catalog, 12 analysts each look up "which event/table should I use" about 3 times a week, at 25 minutes per lookup (asking around, reading old dashboards): 12 × 3 × 25 = 900 minutes = 15 hours per week. With a maintained catalog the same lookup takes 4 minutes: 12 × 3 × 4 = 144 minutes = 2.4 hours per week. The catalog saves 12.6 hours weekly — but only while the entries stay true, which is the next section's problem.
Why catalogs go stale, and how to keep one alive
A catalog is a claim about the present, and the present keeps moving: schemas change, events get renamed, owners leave. An entry with no owner and no update mechanism decays silently — within roughly a quarter, enough entries are wrong that analysts stop trusting the catalog and go back to asking around. Use these decision rules:
- If an entry has no named owner, it is not cataloged yet. "The data team" is not an owner; a person or a small named team is.
- If nothing verifies entries on a schedule, assume they are stale. A "last verified" date older than your review cycle (say 90 days) should visibly flag the entry.
- If new assets can appear without passing through the catalog, the catalog will drift. Route newly seen tables or events through a review step before they become official definitions.
A rollout that respects those rules, in five steps:
- Inventory before you enrich. List the assets people actually query — tables, events, metrics. Start with the 20% that appear in most dashboards, not all 5,000 tables.
- Assign an owner per entry. The owner maintains the definition, not necessarily the pipeline.
- Write the definition and the caveats. One plain-language sentence plus known traps, like the iOS verification caveat above.
- Give entries a status lifecycle. Proposed → in development → in review → active → deprecated. Deprecated entries stay searchable with a pointer to the replacement, so old dashboards can be traced.
- Automate what machines know; schedule what only humans know. Harvest schemas and freshness automatically; put ownership and definition review on a recurring calendar. Name things consistently as you go — see data taxonomy — and treat the plan for new events as a catalog-from-day-one exercise (tracking plan).
The analytics-side catalog in TapCub
Take the signup example above: three teams, three counts, no written definitions. In TapCub the same discipline lives in the tracking plan, the event definitions and the property dictionary, which together catalog one domain: the events and properties in your analytics project. TapCub is not an enterprise data catalog. It does not inventory warehouse tables, BI reports or lineage across systems; for those you still need a warehouse-side catalog such as the AWS Glue Data Catalog.
- Write the entry before the code. In All-platform analytics › Data › Tracking plan, define
signup_completedwith its display name, category, platforms, trigger ("after email verification succeeds"), properties (method: string, required;plan: string, optional), owner and status. The plan holds events you intend to collect, including ones not sent yet, and its statuses are Proposed, Developing, Accepting, Live and Deprecated: the lifecycle from the previous section, built in. Marketing's metric becomes a separate row,signup_form_submitted, with its own owner. - Let real data check the entry. Turn on Live verification (24 h) and the plan adds two columns: hits in the last 24 hours and required-property completeness. If
signup_completedarrives on iOS withoutmethod, completeness drops here instead of in a quarterly audit. Low completeness is a warning; it does not block the event. - Make strays visible. A
track()name nobody registered shows up as Unplanned in All-platform analytics › Data › Event inbox, whose Definitions tab is the catalog of events that already exist. Someone decides whether it gets approved, renamed in code or set to Blocked; until then it is stored like any other event, so the review is a habit, not a gate. - Record what each field means and whether it is sensitive. All-platform analytics › Data › Property dictionary keeps each key's type and its sensitivity (direct identifier, indirect identifier or non-personal), plus optional intake rules. All-platform analytics › Data › Data quality answers the "can I trust it" question for the pipeline: ingestion warnings, collector rejects, high-cardinality properties and possibly stale rules.
- Schedule the human part. The plan has an owner field but no review date, so put the owner review on a calendar, for example every 90 days. The 24-hour hit count tells you an event still arrives; only the owner can say its definition is still right.
The thesis holds at this scale too: the tracking plan works as a catalog only while every entry has an owner and a status someone keeps current. Leave it for a quarter and the event inbox fills with unplanned names nobody reviewed, the same stale list in a smaller domain.
See also: What Is Data Discovery? Four Steps, an Example and Common Mistakes.
Sources
- AWS, What is a Data Catalog? — definition and the business / technical / operational metadata grouping. Checked 2026-11-15.
- IBM, What is a data catalog? — inventory definition and catalog vs. data dictionary. Checked 2026-11-15.
- AWS Glue documentation, AWS Glue concepts — the Glue Data Catalog as a persistent metadata store filled by crawlers. Checked 2026-11-15.
- Collibra, Data Catalog — enterprise catalog scope: definitions, owners, classifications, relationships and policies across systems. Checked 2026-11-15.
Was this explanation clear?
We use your answers to fix and expand this page.
Couldn’t send. Please try again in a minute.
Thanks for the feedback. We’ll use it to improve this page.You’ve already sent feedback on this page. Thanks.

Data Engineer
Writes about event design, data quality, and collection workflows, with verifiable implementation methods for reliable analytics.