I Used an AI Agent to Review This dbt PR: Here's What It Found
The code looked reasonable. Column-level lineage revealed 4 breaking downstream changes before merge.
OVERVIEW
A normal PR diff shows you what changed in the code. It doesn't show you what changed in the data flow: which columns were renamed, which downstream dashboards will break, or whether it's actually safe to merge.
This demo shows a PR review on mart_patient_360, a core healthcare analytics model being refactored to fix a Cartesian explosion, mask SSN for HIPAA compliance, and add financial metrics. The PR touches lineage, sensitive fields, and grain, making it exactly the kind of change where a code review alone isn't enough.
Altimate Code runs a column-level lineage diff between the main branch and the PR branch. It identifies 17 added columns, 3 renames, and 2 removals, along with the downstream blast radius for each change. The most important finding: the primary key (patient_id) was completely missing from the old model and is now added. The most dangerous: renaming full_name to patient_name will silently break every BI dashboard that references that column. The agent surfaces both, along with the grain change from multi-row to one-row-per-patient, before anyone approves the merge.
WHAT YOU'LL LEARN
- →How to run a column-level lineage diff between branches of a dbt project
- →Why code review alone is insufficient for critical analytics models
- →How to assess blast radius on downstream dashboards before merging
- →How to catch breaking changes that look intentional but aren't safe
KEY POINTS
- Compares column lineage between main and PR branch of mart_patient_360
- Identifies 17 added columns, 3 renames, 2 removals with downstream impact notes
- Detects that the primary key (patient_id) was missing from the original model
- Flags SSN masking as a HIPAA fix that may break BI tools using the raw column
- Surfaces grain change: multi-row → one-row-per-patient, with downstream implications
