Inspiration
Everyone ships ML models. Almost nobody notices when an upstream data change quietly rots one in production — a renamed column, a dropped field, a stale table — until the predictions are already wrong and the money is already lost. The dangerous failures are the ones that never throw an error.
What it does
When an upstream dataset changes, Sentinel:
- Traces DataHub's lineage downstream to every production ML model and feature table that depends on it.
- Asks Claude one question per model: will this change silently degrade it? It classifies the risk HIGH / MEDIUM / LOW with a plain-language rationale and a concrete fix.
- Writes the verdict back into DataHub — a Sentinel-At-Risk tag plus a documentation note on the model — so the next engineer or agent inherits the warning instead of rediscovering the outage.
How we built it
DataHub (self-hosted quickstart) as the metadata and lineage backbone. datahub.sdk for the read side (get_lineage, get_lineage_paths_between) and aspect-level writes for the write-back (global tag + editable model properties). Claude (Sonnet 5 / Opus 4.8, via OpenRouter) as the reasoning engine. Packaged as a reusable DataHub Skill (skill/datahub-ml-sentinel) — a sixth skill alongside search / lineage / enrich / quality.
Challenges we ran into
Making the write-back persist without re-emitting DataHub's computed aspects (versionProperties, whose isLatest is read-only) — solved by emitting isolated aspects instead of the whole entity. And modeling a believable ML lineage (datasets to feature table to training job to mlModel) so downstream traversal actually reaches the model.
Accomplishments that we're proud of
The full loop works end-to-end and is verified in DataHub's own UI — the flagged model really shows the tag and the risk note. It's proactive: it predicts future silent degradation, not a lineage viewer. And it ships as an open-source DataHub Skill.
What we learned
DataHub's lineage is generic across entity types, so ML impact analysis is just downstream traversal filtered to ML entities. And the most valuable write-back is the one that lives where the team already works — inside DataHub.
What's next
More change types (freshness gaps, distribution drift), auto-detection via DataHub metadata-change events, and contributing the skill to the datahub-skills registry.
Log in or sign up for Devpost to join the conversation.