Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for DataHub Lineage Sentinel
Inspiration
Production ML failures often originate upstream: an abandoned owner, a failed data-quality signal, or a deprecated feed. Those facts already exist in DataHub, but deployment decisions still depend on manual catalog review.
What it does
Lineage Sentinel queries the official DataHub MCP server for matching production assets, entity metadata, and bounded upstream/downstream lineage. It converts that evidence into clear, review, or blocked decisions, with risk findings in paired JSON and escaped HTML reports.
How we built it
Python 3.11, the official mcp-server-datahub, a strict three-tool read-only allow-list, and a deterministic scoring compiler. A zero-cost synthetic fixture demonstrates missing ownership, PII exposure, failed quality, and deprecated lineage without needing a hosted account.
Challenges
Catalog results are irregular and untrusted. The implementation walks nested responses defensively, bounds queries and asset counts, escapes every rendered catalog value, and fails closed when detail or provenance is missing.
Accomplishments
Reproducible decisions, no model or paid API dependency, no catalog mutations, machine- and human-readable evidence, and five focused tests covering both positive and adversarial cases.
What we learned
DataHub's MCP layer is most valuable when an agent preserves provenance and refuses to treat catalog prose as policy.
What's next
Add authorized live-tenant evaluation, assertion-specific adapters, and signed CI release attestations.
Log in or sign up for Devpost to join the conversation.