Field-Level Signals for Discovering Data Relationships

Joins don’t discover relationships.
They assume relationships already exist.

Signal 1: Null Distribution

· Fields filled in almost every row → behave like identifiers

· Sparse fields → contextual attributes

Null patterns tell you more than column names.

Signal 2: Cardinality (and Why It Saves You from Bugs)

· High cardinality → identifiers

· Low cardinality → states / enums

This is how people accidentally join unrelated status fields.

Signal 3: Inclusion Beats Equality

· Real systems are asymmetric

· One table is usually upstream

· Another is delayed / filtered

Why Brute-Force Comparison Doesn’t Scale

· Millions of values

· Exponential field growth

· Compute explosion

· Feature extraction

· Sampling

· Staged comparison

Practical Checklist

In the final post, I’ll argue why relationship discovery shouldn’t be treated as analysis work at all — but as infrastructure.