Bruin Self-Heal Demo

This project is a local DuckDB sandbox for testing Bruin agent troubleshooting skills. It has demo-seed to load realistic raw tables, then demo-pipeline with normal SQL assets and four intentionally failing branches.

Install the agent skills separately if you want an agent to use them:

bruin ai skills all
bruin init self-heal-demo

What Is What

Load the raw tables first:

Pipeline	Assets	Purpose
self-heal-demo/demo-seed	raw.orders, raw.order_status_history, raw.order_adjustments, raw.fulfillment_events, raw.product_catalog	Seeds DuckDB source tables used by the demo.
self-heal-demo/demo-pipeline	staging.orders, orders.status_snapshot, finance.order_margin, fulfillment.daily_activity, catalog.product_prices	Runs the intentionally failing analytics branches.

The main pipeline uses normal business asset names. Scenario names appear only as tags:

Scenario tag	Assets	Failure symptom
duplicate-investigate	staging.orders, orders.status_snapshot	orders.status_snapshot fails the uniqueness check on order_id.
quality-check-investigate	staging.orders, finance.order_margin	finance.order_margin fails the positive check on net_amount.
freshness-check	fulfillment.daily_activity	fulfillment.daily_activity is missing the latest source activity date.
schema-drift-check	catalog.product_prices	catalog.product_prices fails because the raw catalog schema changed.

Run

From the generated project root, the directory that contains .bruin.yml and the self-heal-demo/ folder:

bruin run self-heal-demo/demo-seed
bruin validate self-heal-demo/demo-pipeline

Validation should pass after seeding because the broken scenarios are runtime data or warehouse issues.

Run one scenario branch at a time:

bruin run --tag duplicate-investigate self-heal-demo/demo-pipeline || true
bruin run --tag quality-check-investigate self-heal-demo/demo-pipeline || true
bruin run --tag freshness-check self-heal-demo/demo-pipeline || true
bruin run --tag schema-drift-check self-heal-demo/demo-pipeline || true

Proof Queries

After running a branch, use the DuckDB connection to inspect the specific bad row, partition, or schema.

bruin query --connection self-heal-demo --query "SELECT order_id, count(*) AS row_count FROM orders.status_snapshot GROUP BY 1 HAVING count(*) > 1;"
bruin query --connection self-heal-demo --query "SELECT order_id, gross_amount, adjustment_amount, net_amount FROM finance.order_margin WHERE net_amount <= 0;"
bruin query --connection self-heal-demo --query "SELECT max(event_date) AS modeled_max_date FROM fulfillment.daily_activity;"
bruin query --connection self-heal-demo --query "SELECT max(CAST(event_timestamp AS DATE)) AS raw_max_date FROM raw.fulfillment_events;"
bruin query --connection self-heal-demo --query "DESCRIBE raw.product_catalog;"

Expected results:

• The duplicate query returns order_id = 1002 with row_count = 2.
• The quality query returns order_id = 1003 with a negative net_amount.
• The modeled fulfillment max date is 2025-01-02, while the raw fulfillment max date is 2025-01-03.
• The catalog description shows listed_price; the failing model still expects the previous pricing field.

Expected Diagnosis And Fix

Scenario tag	Diagnosis	Recommended fix
duplicate-investigate	Status snapshot joins all status history rows, duplicating multi-status orders.	Filter status history to the current status before joining.
quality-check-investigate	Adjustments are grouped by customer, applying another order's refund incorrectly.	Aggregate and join adjustments by order_id, not customer_id.
freshness-check	Fulfillment model excludes latest delivered activity through narrow event filtering.	Include delivered events or align filtering with activity coverage.
schema-drift-check	Product price model references removed unit_price after source rename.	Select listed_price AS price from raw.product_catalog.