The kit of a database schema: tables, with columns that have a name, a type and a key (primary or foreign), and foreign keys drawn from a column to a column of the same type.

A table is a node of type "schema-table" whose data is:

%{
  "name" => "orders",
  "columns" => [
    %{"id" => "c1", "name" => "id", "type" => "uuid", "key" => "pk"},
    %{"id" => "c2", "name" => "customer_id", "type" => "uuid", "key" => "fk"},
    %{"id" => "c3", "name" => "placed_at", "type" => "timestamp", "key" => nil}
  ]
}

Each column has a handle on each side, of the column's family of types (family/1): an edge joins two columns of one family (integer to bigint, varchar(64) to text, but not text to uuid), and takes the sides nearest each other. A column's handles are its "id" (its name, for a column without one), so a foreign key stays when people rename the column. People rename the table and its columns in place, pick a column's type, set its key, and add and remove columns.

A table of another schema than the one shown (the table a foreign key points to, in billing) is a stub: data with "external" => true and its "schema", and only the columns its foreign keys use. It is drawn faint, with a dashed border and its qualified name (billing.invoices), its edges dashed, and nobody edits it, even on a canvas they edit:

%{"name" => "invoices", "schema" => "billing", "external" => true,
  "columns" => [%{"name" => "id", "type" => "uuid", "key" => "pk"}]}

A table may name its "schema" without being a stub; its name shows qualified too.

In Markdown, a schema is a Mermaid entity relationship diagram: each table with its columns (uuid id PK), and a foreign key as a relationship from its table to the table it refers to, named after its column. from_markdown/1 reads it back, and reads ER diagrams people write: a relationship becomes an edge from the column it names (else the table's first column) to the other table's primary key.