Architecture
How Fabric Experiments separates control plane, edge data plane, warehouse analysis, and agent operations.
Fabric Experiments is split into four planes, with Fabric Platform providing the governed mutation spine and Fabric Harness providing workload execution.
Control plane
The control plane owns experiment lifecycle, project/property/feature configuration, immutable environment publications, authorization, audit events, manifests, manifest keys, aggregate queries, and internal worker APIs. Mutations pass through the canonical Platform action path: policy evaluation → state validation → handler → events.
Edge data plane
The edge data plane serves low-latency manifests and accepts high-volume exposure/conversion events without routing every beacon through the control plane. Tenant-scoped ingestion is rate-limited and validates each event before enqueueing it. Cloudflare Queues provides durable acknowledgement, batching, retries, and a dead-letter queue; deterministic R2 keys make redelivery idempotent.
Warehouse analysis
The Asset Bundle provisions a Unity Catalog schema, managed ingestion Volume, serverless SQL warehouse, Lakeflow Auto Loader pipeline, and refresh Job. Azure deployments use an idempotent, bucket-scoped R2 sync task before Auto Loader; Databricks on AWS can use a native Unity Catalog R2 external location instead. The local DuckDB warehouse adapter supports tests and examples.
Agent operations
Harness/Temporal workers run SRM watcher and analyst workflows. Databricks experiment bindings cover SQL, notebooks, Jobs, Lakeflow, models, and Feature Serving in online, offline, or shadow modes. These workflows operate through governed API/action surfaces rather than mutating storage directly.