Build a one-way pipeline from PostgreSQL into an air-gapped, governed data zone — where access is controlled at the database, table, and column level.
The New Problem: AI + Production Databases
Modern stacks are quietly introducing a dangerous pattern:
AI Agent → PostgreSQL
Even when wrapped with APIs, proxies, or query layers, the reality is:
- The agent still reaches production
- It still executes queries
- It still operates inside your blast radius
This is fundamentally unsafe.
Not because AI is malicious — but because it is non-deterministic.
The Design Principle
pg-cdc is built around a simple rule:
AI should never have network access to production databases.
Instead of securing access…
We remove it entirely.
What pg-cdc Actually Is
pg-cdc is:
- A PostgreSQL logical replication (WAL) consumer
- A Parquet writer with compaction
- A governance-aware ingestion layer
- A security boundary between production and consumers
The Air Gap
The most important property:
pg-cdc creates a physical and logical air gap
Properties
- No return path
- WAL is unidirectional
- Storage is append/immutable
- No database credentials
- Consumers use IAM
- No connection strings
- No shared execution environment
- No queries run on PostgreSQL
- Governed surface only
- Access mediated via catalog + tags
- Auditable: Every read can be tracked
If You’re Building This Kind of System
Think in layers:
- Production system (OLTP) → isolated
- CDC boundary (pg-cdc) → one-way
- Governed storage → immutable + tagged
- Query layer → controlled access
- AI / analytics → consumers
I’d love to hear — these are exactly the problems we’re solving.
Repository: GitHub repo
Connect:LinkedIn