DataMeshOps: Autonomous Data Remediation & ML Governance

Autonomous agents that fix broken data pipelines and quarantine compromised ML models by reading and writing back to the DataHub metadata graph via MCP.

Comment

Inspiration

Data engineering teams spend 40% of their time manually triaging broken pipelines and silent ML failures. DataHub's MCP Server unlocks the ability to build agents that don't just read metadata — they act on it and write knowledge back for the next engineer or agent.

What it does

DataMeshOps has two autonomous agents built on a shared datameshops-core library: DataMeshOps-Healer — Detects SLA breaches and data quality failures, traces 3-hop upstream lineage via DataHub MCP, synthesizes a non-destructive SQL remediation patch, then writes remediation metadata back to DataHub: #auto-remediated tags, structured audit properties, a Markdown post-mortem runbook, and resolved incident status. DataMeshOps-Guard — Audits production ML models by tracing feature store lineage through DataHub, detects target leakage and upstream schema drift, then quarantines the compromised model: #ml-quarantined tags, model_health_score property, diagnostic runbook, and a CRITICAL blocking incident — preventing downstream CI/CD from deploying a broken model.

How I built it

  • datameshops-core: Async MCP transport manager with exponential backoff, multi-hop graph traverser, provider-agnostic LLM adapter (OpenAI / Gemini / Anthropic), and atomic mutation engine
  • DataHub integration: mcp-server-datahub via stdio/HTTP, DataHub GMS GraphQL/REST endpoints
  • CLI: Typer + Rich for staged, color-coded terminal output with real-time spinners
  • Zero-cost judge testing: deterministic mock fallback — no API key or Docker needed to run

DataHub MCP Tools Used

search, get_lineage, list_schema_fields, get_dataset_queries, add_tags, add_structured_properties, update_description, raise_incident, update_incident_status

Challenges

  • Getting TOOLS_IS_MUTATION_ENABLED mutation flow to work end-to-end with proper fallback for offline judge evaluation
  • Designing the agent loop so metadata written back to DataHub is immediately readable by subsequent agents (knowledge inheritance)

What I learned

DataHub's MCP Server makes it possible to build agents that don't just observe data problems — they resolve them and document the resolution for the next human or agent. The graph mutation capability is the real unlock.

Open-Source Contribution

Includes DataHubBatchMutatorSkill (datameshops_core/batch_mutator.py) — a reusable high-throughput batch metadata mutation skill ready for PR to datahub-project/datahub-skills.

Built With

  • asyncio
  • datahub
  • geminiapi
  • graphql
  • mcp
  • openaiapi
  • python
  • rich
  • typer
Share this project:

Updates