RepairPilot

The DataHub-powered agent that blocks dangerous dbt changes, proves a real repair, gates release on owner approval, and writes reusable knowledge back to DataHub.

Comment

Inspiration

Most incident agents stop after explaining what broke and sending a message. A data platform team still has to assess risk, edit production code, prove the repair, collect authority, and preserve what was learned. RepairPilot closes that gap while keeping the probabilistic model behind deterministic safety rails.

What it does

The demo starts with Northstar Commerce renaming gross_amount to gross_revenue. DataHub MCP reveals that the field is a Tier-1 governed financial metric owned by Revenue Analytics, with three downstream dbt models, an executive dashboard, three stored usage queries, and active quality checks. A versioned policy scores it HIGH/100 and blocks the release.

DeepSeek V4 Flash proposes a compatibility alias, a controlled downstream migration, and a new schema test. RepairPilot never executes generated shell or SQL. It applies fixed, reviewed templates only to three model-editable allowlisted paths in a detached Git worktree and per-run Postgres schemas; the executor adds the fourth, reviewable migration note. The breaking build must fail; the repaired dbt build --select stg_orders+ must then pass. Only after 14 tests pass does the owner approval control unlock.

After approval, judges can inspect the real Git patch and GitHub PR. RepairPilot also writes an Incident Document, reusable safe-rename Runbook, RepairPilotVerified tag, and successful gross_revenue Assertion back to DataHub so the next human or agent inherits the evidence.

How we built it

  • DataHub OSS 1.6.0 and official DataHub MCP Server 0.6.0
  • PostgreSQL 16, dbt Core 1.10, and dbt-postgres
  • FastAPI/Pydantic state machine and deterministic policy engine
  • DeepSeek V4 Flash with a strict JSON repair contract
  • React/Vite Repair Control Room
  • Git worktrees, GitHub Actions, and a canonical review PR
  • Caddy HTTPS and hardened systemd services on Ubuntu 24.04 ARM64
  • Remotion for the under-three-minute product video

Challenges

The hard part was proving that every impressive-looking step was real without giving the model unsafe authority. DataHub MCP output had to be traced into the UI, DataHub OSS version differences required fail-closed mutation validation, and dbt had to demonstrate both the original failure and the repaired success in clean isolation. Public hosting also required loopback-only data services, credential isolation, a global run lock, and repeatable reset behavior.

Accomplishments

  • complete DataHub read/action/write-back loop rather than metadata chat;
  • immutable receipt connecting DataHub URNs, policy version, Git commits, Patch SHA-256, dbt invocation, approval, PR, and writeback URNs;
  • three consecutive Live runs with the same repair hash and idempotent Runbook;
  • real failing and passing dbt builds with 14 tests;
  • 83%+ Python coverage, public HTTPS deployment, and no known flaky tests;
  • sample artifacts that judges can inspect without running the service.

What we learned

Context makes an agent useful, but authority boundaries make it trustworthy. DataHub is most valuable here not merely as a catalog lookup: it supplies the governance facts that determine risk, then becomes the institutional memory for the verified outcome.

What's next

Add signed enterprise identity for approvals, policy-as-code packages per Domain, automatic consumer migration windows, and optional Airflow/Dagster pause and rerun adapters. The fixed demo templates would become repository-specific repair skills reviewed by each data platform team.

Data disclosure

All Northstar Commerce rows, people/groups, dashboard, and query metadata are deterministic synthetic demonstration data. No personal, customer, or production data is used. DataHub, dbt, Postgres, MCP, DeepSeek, Git, and GitHub executions are real.

Source/provenance disclosure

The project was newly created during the July 6–August 10, 2026 submission period. It contains no pre-existing proprietary application code. It uses the open-source packages and images declared in the repository and was developed with AI coding assistance.

Built With

Share this project:

Updates