Gantry¶
Gantry is a governed execution boundary for agent-generated SQL and data jobs. It decides whether a statement may run, bounds it while it runs, and decides afterwards whether the result may be believed.
pip install data-gantry
The distribution is data-gantry; the import is gantry.
Where to start¶
| SQL | the governed query path, and the agent-facing tool |
| Policies | what a policy can express, and who enforces it |
| Materialization | letting an agent build a table, and checking it |
| Flink backends | batch and stream jobs, and what each guarantees |
| Adapters | the interfaces a new backend implements |
| API reference | every public symbol, generated from source |
Seven runnable examples live in
examples/, indexed
by task. The quickest needs nothing but a Python environment.
The idea in one paragraph¶
An engine answers "did the statement run". That is not the same question as "is
the answer usable" — a CREATE TABLE AS whose join matched nothing succeeds,
and a streaming job that has restarted forty times is RUNNING. Gantry answers
the second question, and refuses the statement outright when a policy says it
should never have been asked.