DataHub hackathon · self-healing data pipelines

DAGtor

An autonomous SRE agent that builds your SQL data pipelines from a chat prompt, heals them when they break at runtime — diagnosing root cause against DataHub metadata — and learns from every fix so recurrences repair instantly.

Two pillars

A builder and a doctor in one agent

Conversational pipeline builder

Describe an analysis in plain English. DAGtor grounds it in real DataHub tables, columns and PII tags, then generates a runnable DAG — two loads → join → group → filter — with PII columns pre-masked. Review it on the canvas, then deploy.

Gemini function-callingDataHub-grounded SQLtemplate fallback offline

Self-healing & self-learning

When a stage fails at runtime, DAGtor intercepts it, diagnoses the root cause in DataHub's metadata graph, rewrites the SQL (masking PII), and re-runs. Every validated fix is saved — the next identical failure heals instantly from memory.

🔴→🟡→🟢 live stateshuman-in-the-loop approvals⚡ instant recall
Under the hood

One process, standard APIs

Full architecture
Browser
Agent chat · React Flow canvas · Activity / Results / Healing views
WebSocket + HTTP
Next.js server (one Node process)
API routes
/api/chat · /api/status · /api/experience
WebSocket bus
live state fan-out
Orchestrator
server-side state machine — drives run loop, node states & the heal/approve cycle
Gemini
build + diagnose (function calling)
DataHub MCP
schema fields · lineage · PII tags
SQL executor
real SQL over fiction-retail (SQLite)
Experience Store
learned fixes (SQLite memory)
Built with

The stack

DataHub MCP

The metadata layer — schema fields, lineage and PII/governance tags — reached over the Model Context Protocol.

Gemini

Classifies intent, builds pipelines, and runs the function-calling diagnosis loop over the DataHub tools.

Next.js + React Flow

App Router UI with a live DAG canvas; a custom Node server streams real-time state over WebSockets.

SQLite + PostgreSQL

One database, chosen by env: SQLite locally, PostgreSQL for remote deploys. Holds the source tables the pipeline runs on plus the Experience Store of learned fixes.

Built on DataHub

Grounded in the metadata graph

The challenge asks entrants to build on the open-source DataHub platform plus at least one of four agent building blocks. DAGtor points at DataHub Core (OSS) through the official MCP server — a config-only swap, with a protocol-compatible local server for offline dev — and uses three of the four building blocks, so every diagnosis is grounded in real metadata, never guessed.

MCP Server

Speaks the Model Context Protocol to the official mcp-server-datahub against a live DataHub Core (OSS) graph — or a protocol-compatible bundled server for offline dev. The same tool calls either way.

Agent Context Kit

Equips Gemini with grounded DataHub context — schemas, lineage and governance tags — so the agent reasons over the real metadata graph instead of hallucinating.

DataHub Skills

The official tool registry (npx skills add datahub-project/datahub-skills) gives the agent explicit tools — get_lineage(), list_schema_fields() — it calls while healing.

Analytics Agent

Not used in this build.

Self-healing pipelines that learn

Ask DAGtor for a pipeline and deploy it. When a stage drifts at runtime — a renamed column, a dropped upstream — it diagnoses the root cause against DataHub, repairs the SQL, and remembers the fix so it heals instantly next time.

Launch app