DataHub-powered ML change safety

Stop silent data changes before they break production models.

LineageGuard traces a schema event through DataHub's context graph, scores the true ML blast radius, proposes the smallest safe repair, proves it with dbt, and preserves the decision for the next engineer or agent.

Watch 2:19 demo

This public judge experience replays a captured, reproducible run against real DataHub OSS 1.6.0 and the official DataHub MCP server 0.6.0. It does not claim to host the full DataHub stack at the edge.

One bounded agent loop

Evidence before automation

Every step below corresponds to evidence preserved in the public repository and narrated demo.

01 / EVENT

Breaking rename

customer_age becomes age_years without compatibility.

02 / CONTEXT

Trace DataHub

Read ownership, PII, Tier1, and three-hop ML lineage through MCP.

03 / DECIDE

Score blast radius

Deterministic policy exposes every point behind the 90/100 decision.

04 / PROVE

Validate repair

Restore the stable contract and run the real dbt build in a temporary copy.

05 / REMEMBER

Write back

Publish the decision as durable DataHub context for the next responder.

RiskCritical · 90

RepairCompatibility alias

dbt proof9 / 9 pass

DecisionWritten back

Captured from the verified run

Inspect the real graph

These are direct screenshots from the local DataHub stack used in the submission—not generated mockups.

DataHub dataset showing schema, ownership, PII and Tier1 tags — **Governed source asset**Schema, technical owner, PII, and Tier1 context.

DataHub lineage graph connecting the changed dataset to ML entities — **ML blast radius**Dataset → feature → model → production deployment.

LineageGuard critical decision document written back into DataHub — **Durable decision writeback**The incident outcome becomes reusable catalog context.

Reproducible, bounded, honest

Verification ledger

What is proven

Real DataHub OSS 1.6.0 graph and official MCP server 0.6.0
Explainable 90/100 critical risk decision
All nine dbt nodes and tests pass after the bounded repair
Decision document persisted as urn:li:document:lineageguard-datahub-31e4e1b6a72c9344

Safety boundary

MCP access is read-only; mutation tools are disabled
Patch must match exactly once and is applied only in a temporary copy
GitHub output, when enabled, is always a draft PR
Zero-credential fixture and locked Docker path are independently reproducible