Integration · Warehouse
Snowflake
Snowflake is where the warehouse-side reconciliation happens. The customer's data team can reproduce every doubly-robust lift readout we bill against by running offpolicy.estimators.dr(...) over the same logged assignment + outcome rows in their warehouse. The math is auditable.
The contract
What flows in
- Optional: Snowflake share / external table for assignment + outcome rows we ingest for the audit / 90-day replay
- Schema mapping from your warehouse columns to the offpolicy.py canonical schema
What flows out
- metapolicy_lift_readout table written to your warehouse via Snowpipe or direct INSERT
- policy_snapshot_id traceable from each readout row back to the deployed bandit policy at decision time
Setup
- 01Create a Snowflake share or grant the Metapolicy service role read access to your assignment + outcome tables
- 02Map column names to the offpolicy.py canonical schema (the dashboard has a UI to make this trivial)
- 03Start the 90-day replay; it runs against your warehouse directly — nothing leaves your perimeter for the audit deliverable
- 04Optional: enable Snowpipe push so lift readouts land in your warehouse within seconds of OPE-worker completion
Warehouse-side reproducibility is the trust signal
The single hardest objection from a CFO doing diligence on outcome pricing is: how do I verify you didn't fudge the lift number?
Snowflake (or BigQuery, or Redshift) is the answer. The DR estimator is open-source on PyPI. The propensity is logged in the same row as the assignment. Your warehouse already holds both. Run the math; compare. The same number we billed will appear.
Wire Snowflake in a 30-minute call.
We pair with your engineer, ship the integration live, and run the first decision on your stack.