Skip to main content

Catalog Governance

One catalog covers every database, API and GraphQL service you connect, and every table your team or your automations create. People and agents curate it together, lineage is recorded from the work that actually runs, every change can be rolled back, and what it says about a table travels with every answer your agent gives.

Coverage

What the catalog holds

A table belongs to the connection that sees it, so the same table reached through two connections is two catalog entries, each governed by who may use that connection. Nothing is imported or synchronized from elsewhere: the catalog reads what your connections see and records what the platform creates.

A scan never touches curation. When a table disappears from a source, its entry is marked removed and its descriptions stay readable; if the table returns, the mark clears and the curation is still there. A source that cannot be read in full is recorded as unread, with the reason, and nothing under it is marked removed.

SourceCataloged asDetailWhen
Databases and warehousesEvery table and view each connected catalog exposesEach column, in order, with its type and whether it allows nullsOn a regular scan of each connection, and whenever an admin asks
REST APIsEvery operation in the connection’s API descriptionNo columns; the connection’s API description is kept beside it as a versioned documentOn the same scan; a new version is kept only when the description changes
GraphQL servicesEvery query and mutation in the schemaEach argument, with its typeOn the same scan
Spreadsheets and files registered as tablesThe table the registration createsThe columns the registration declaresThe moment the table is registered, before the request returns
Webhook, form and mailbox sourcesThe table each source fillsThe event columns every source table carriesThe moment the source is created

What can be written on each entry

A table and each of its columns
Description, tags, glossary terms, owners, domain, custom and structured properties, documentation links
A table only
Curated queries, context documents (markdown notes with a title and a category), open and resolved data incidents
Glossary terms and folders, tags, domains, data products
Description, owners, custom properties, documentation links, and the list of tables and columns that carry them

Admins choose how often each connection is scanned and which catalogs, schemas or tables to leave out, under Admin, Catalog, where each connection’s last scan is listed with what it wrote, what it marked removed, and anything it could not read. Files in object storage and the tools of connected MCP servers are reached through their own connections and are not cataloged as tables.

Curation

People and agents write to the same catalog

Most catalog curation used to be a backlog: descriptions nobody had time to write, glossary terms agreed in a meeting and never attached to a column. Here the agent does much of the first draft. When it learns during a conversation that a column holds gross amounts or that a table stopped receiving rows in March, it captures that as an insight. A reviewer approves it, merges it with others, or rejects it, and only then does it reach the catalog.

A steward who would rather type the definition opens the table under Knowledge, Catalog and edits it in place. Tags are picked from the vocabulary or typed as new; glossary terms and domains are picked by name. Both paths land in the same catalog, pass the same permission checks, and leave the same record.

A person in the portal

  1. 1Opens the table in Knowledge, Catalog
  2. 2Edits the description, tags, terms, owners or domain in place
  3. 3The write lands at once and is recorded with their name

An agent in a conversation

  1. 1Captures what it learned as an insight
  2. 2A reviewer approves, merges or rejects it in the review queue
  3. 3The approved change is written as one changeset that rolls back as a unit

One catalog, one set of rules

  • Reading takes the role’s read grant; writing takes the curation grant
  • Only datasets on connections the role reaches are shown or written
  • Every write is audited and kept in the field’s history

A definition still in use cannot be deleted: a glossary folder with terms in it, or a term, tag or domain that a table or column carries, is refused with a list of what still uses it. How insights move from a conversation to the catalog is covered on Knowledge Capture & Application.

Lineage

Lineage recorded from what runs

Hand-drawn lineage goes stale the week after it is drawn. Plexara records an edge when the platform itself moves data, and names what wrote it:

  • a SQL statement that writes a table from what it reads (create table as select, insert from a select, merge, or a view), from each table read to the table written;
  • a successful automation run, from every table its queries read to every file it exported;
  • registering a table over a file, from the file to the table.

A table’s page shows its lineage three steps upstream and three downstream, each step a link. Steps on connections the reader cannot use are left out.

Recorded lineage, ACME demo deployment

  • Webhook table

    scratch.uploads.webhook_pos_sales

  • Warehouse table

    warehouse.public.stores

  • Warehouse table

    warehouse.public.regions

Written by

acme-live-sales-dashboard

an automation, every 15 minutes

Dashboard in the portal

ACME Live Sales

Revenue, sales by hour, regions and tender for the last 24 hours

Nobody drew these edges. Each successful run of the automation records one edge from every table its queries read to the file it wrote, and names the automation as the writer. Open the stores table in the catalog and the dashboard is listed downstream; open the dashboard and the three tables are listed upstream.

Descriptions follow the lineage down

A derived table usually repeats columns from the tables it was built from, and usually arrives with none of their descriptions. When a column has no description or glossary term of its own, Plexara shows the one on the same-named column of the nearest upstream table that has one, up to two steps away, and marks where it came from. Describe customer_email once on the source table and every table built from it explains the column too.

Tags are never borrowed, because a tag such as pii is a statement about the column itself. Nothing is copied downstream either: the borrowed text is read from the source each time, so correcting it upstream corrects it everywhere.

Lineage runs table to table, and covers the work done on the platform and the SQL that writes through it. Pipelines and BI tools that run elsewhere are outside it.

History

Every change kept, any change rolled back

When agents write to the catalog, a reviewer needs more than an audit log. They need to see what a field said before, who changed it, and to put it back. Every field on every entry keeps its own history: the new value, the value it replaced, who wrote it and when. Roll back restores the previous value and records that as a change too.

History of one field

transactions.total, description

  1. v3

    A steward, in the portal

    Oct 22, 14:08

    Gross sale total including tax. Returns are separate rows with a negative total.

    Replaced: v2

    Roll back
  2. v2

    An agent, applying a reviewed insight

    Oct 19, 09:41

    Gross sale total including tax.

    Replaced: v1

    Written over since: a roll back here is refused, naming the newer write
  3. v1

    An analyst, in the portal

    Oct 14, 06:05

    Sale amount.

    Replaced: empty

    Written over since: a roll back here is refused, naming the newer write
An example of how a field’s history reads. Each field keeps its own history, so rolling back a description leaves the tags and the owners beside it alone. A roll back is itself recorded as a change, and an edit made from a value someone has since replaced is refused and asks the editor to reload rather than overwriting their colleague.

An approved insight that wrote several fields at once rolls back as one changeset, and the insight returns to the review queue with a note of what was tried. A roll back that would discard a newer edit is refused and names that edit, so undoing one person’s change never silently removes another’s.

For the agent

The catalog travels with every answer

A catalog that people browse is useful. A catalog the agent reads before it writes a query changes the answer. One call returns a table’s whole record: its columns, everything written on it, its lineage, and whether it can be queried right now and where.

One table, one call

What the agent calls. This is the exchange the agent has with Plexara on your behalf, shown for the technical reader. You ask in plain language; you never type any of it.

fetch(reference: "mcp:dataset:(trino,acme,warehouse.public.stores)")

{
  "name": "stores",
  "type": "table",
  "columns": [
    { "name": "store_id",       "data_type": "integer",      "nullable": false },
    { "name": "square_footage", "data_type": "integer",      "nullable": true  },
    { "name": "is_active",      "data_type": "boolean",      "nullable": true  }
    // 9 more columns
  ],
  "aspects": {
    "description": "Master list of all ACME Corp retail locations...",
    "tags": ["DimensionTable", "CoreEntity"],
    "domain": "mcp:domain:da265467-...",
    "context_documents": [
      { "title": "Store Classification & Regional Operations",
        "category": "business_rules" }
    ]
  },
  "downstream": [
    { "target": "mcp:asset:1057d934-...",
      "writer_kind": "script",
      "writer_ref": "mcp:script:c882969a-...",
      "depth": 1 }
  ],
  "query_availability": {
    "available": true,
    "query_table": "warehouse.public.stores",
    "connection": "acme"
  }
}
The stores table on the ACME demo, shortened. The downstream edge is the live sales dashboard from the lineage figure above. The agent learns that the table is queryable, what a steward wrote about store size bands, and what depends on it, before it writes a line of SQL.
With every query result
A query result arrives with the description, owners, tags, domain and glossary terms of the tables it read, and the column notes for the columns the statement used. Open data incidents come with it, so the agent knows a table is under investigation before it reports from it.
In one search
Search ranks catalog tables by meaning as well as by name, beside glossary terms, knowledge pages, saved reports and what your team has captured. A search about a specific table widens one step along its lineage, so what is known about its sources surfaces too.
Under the reader’s access
Fetch returns only what the role may reach. A glossary term’s list of tables counts the ones on connections the reader cannot use instead of naming them.
As text a model can trust
Catalog text is cleaned before it reaches a model, and instructions planted in a description are stripped and logged.

In production

Sized for a real estate of systems

The client deployments behind these figures run the same catalog every customer gets, with their own systems behind it.

Thousands
of warehouse tables cataloged in one media deployment
200+
glossary terms in the same deployment, defining what its tables mean
4,000+
ERP objects cataloged beside point-of-sale data in a retail deployment

Where it fits

Your first catalog, or the one beside your program

For a company whose catalog will mostly be read by AI agents, this catalog does the job a first catalog purchase would: a glossary, ownership, lineage of the work that runs, and history, kept where the agents query and the automations run.

An enterprise governance program has needs it does not cover: column-level lineage extracted from dozens of ETL and BI tools, data quality monitoring and profiling, certification workflows, and privacy scanning at scale. Those programs are well served by dedicated catalogs, and the data catalog comparison sets out where each fits. The persona model that governs every read and write is on Governance, and lesson 211, catalog governance in the portal, walks the screens.

Common questions

Catalog FAQ

Every database and warehouse table and view your connections can see, every operation of a connected REST API, every query and mutation of a connected GraphQL service, and every table your team or your automations create by registering a spreadsheet or file or by adding a webhook, form or mailbox source. Connections are scanned on a regular cycle and whenever an admin asks; created tables are cataloged the moment they exist. Files in object storage and the tools of connected MCP servers are reached through their own connections rather than cataloged as tables.

Learn more: Integrations

Anyone whose role holds the curation permission can edit the tables on connections that role reaches, in the portal or through their agent, and every edit is audited. Agent-captured knowledge goes through the review queue before it is written. Every field keeps its own history with the previous value, who changed it and when, and Roll back restores it. A roll back that would discard a newer edit is refused and names that edit.

Learn more: Governance

From the work the platform runs. A SQL statement that writes one table from others, a successful automation run that exports a file, and the registration of a table over a file each record an edge that names what wrote it. A table shows its lineage three steps upstream and downstream. Lineage runs between tables, not columns, and pipelines or BI tools that run outside the platform are not part of it.

It records data incidents a person or an agent raises against a table, and those incidents travel with every query result that reads the table, so the agent knows a table is under investigation. It does not run quality checks, profile data, or keep certification badges. Companies that need those alongside an AI program usually run a dedicated quality tool or catalog for them.

Learn more: Data catalogs compared

Before it queries, the agent can fetch a table’s whole record in one call: columns, descriptions, tags, glossary terms, owners, context documents, lineage, and whether it can be queried and where. After it queries, every result arrives with the descriptions and column notes of what it read. A column with no description borrows the same-named column’s from up to two steps upstream, marked with where it came from.

Next

Governance

The persona model behind every catalog read and write, enforced at the point of execution.