Warehouse aggregation
Exposure and conversion data lands in R2 and is aggregated through Databricks or a local adapter.
High-volume telemetry belongs in the data plane and warehouse, not the control-plane event store.
Flow
The API exposes experiments.aggregate, which returns typed per-variant rows
plus time-to-convert, depth, segment, and CUPED analytics. Cloudflare uses the
Databricks Statement Execution REST API with OAuth M2M token refresh; Node
deployments can use the lazy-loaded SQL driver.
Production refresh runs every 15 minutes. On Azure Databricks, a bucket-scoped, read-only R2 credential copies new immutable objects into a governed Unity Catalog Volume before Lakeflow processes them. An AWS-native external-location design is possible, but AWS and GCP are not supported targets until the full compatibility certification runs there.
For Databricks deployments, install:
pnpm add -F @fabricorg/experiments-api @databricks/sql