Fabric Experiments

Fabric Experiments is a Databricks adoption accelerator and TypeScript quality engineering platform. Databricks remains the execution, data, AI, lineage, and governance foundation. Fabric connects four workflows that are normally split across unrelated products:

1. Experiment — design, deliver, and analyze governed A/B tests.
2. Validate — test Databricks SQL, Delta, Lakeflow, Jobs, notebooks, dbt, Unity Catalog, Volumes, Apps, and Lakebase with Gherkin and TypeScript, locally and against a real workspace.
3. Evaluate — use managed MLflow as the native evaluation system of record, link its run evidence into Quality Center, or run application-side TypeScript checks against Databricks Model Serving.
4. Govern — publish evidence to Quality Center, enforce production gates, retain tenant-scoped audit history, and promote configuration safely.

It combines a Git-first CLI workflow, a hosted control plane, signed edge manifests, lightweight SDKs, warehouse-backed aggregation, and Fabric Harness-powered agents for operational guardrails and analyst readouts.

This is not just an A/B testing tool. Databricks teams can automate workload behavior, contracts, permissions, performance, recovery, and release readiness. Focused unit, browser, load, and security tools can publish their evidence into the same Quality Center and participate in production gates.

You do not need Python. Greenfield teams write step definitions, target adapters, and support code in TypeScript; Python, SQL, R, and Scala notebooks, pipelines, and models can still be the workloads under test.

Fabric does not recreate managed MLflow, Unity Catalog, Model Serving, AI Gateway, SQL, Delta, or Lakeflow. It drives and observes those native services, retains links to their authoritative evidence, and adds cross-workload release automation around them.

New here? Pick your path

Test a Databricks workload

Install the published TypeScript packages, run a Gherkin scenario locally with DuckDB, move the same intent to a real workspace, and publish evidence.

Explore first-class target packs

Run the shipped SQL, Delta, Lakeflow, Jobs, Unity Catalog, Apps, Lakebase, and storage packs; diagnose configuration and publish Quality evidence.

Evaluate AI applications

Build reproducible TypeScript evaluation suites with versioned datasets, deterministic Databricks model judges, prompt hashes, OpenInference traces, token budgets, and warehouse evidence.

Run experiments in Studio

Use the lifecycle Overview, six-stage Pipeline, guided experiment builder, results, Quality Center, governance, and audit history in one organization.

Secure an enterprise rollout

Review available identity, tenant, evidence, reliability, privacy, recovery, and Databricks controls together with current assurance boundaries.

60-second A/B testing quickstart

Install fx, write one YAML file, and run an exposure → conversion → report cycle on your laptop. No account or infrastructure required.

Compare Mojito and Behave

Evaluate behavior and target coverage using explicit shipped, live-gated, certifying, different-by-design, and unsupported statuses.

Building something more advanced?

All YAML fields

Complete schema reference: domOps, the Trigger DSL, sharedJs/sharedCss, recipeSampleRate, metadata, and the holdback / sampleRate / weight triad explained with examples.

Pluggable SDK internals

Custom storage adapters (cookie / localStorage / session / your own), decision adapters for bandits or feature-flag systems, exclusion predicates for bots and consent gating, and separate hooks for variant-code failures.

Production deployment

Provision Cloudflare, Databricks Apps, Lakebase, Databricks SQL, and Temporal for a governed, multi-tenant install with audit forwarding and live-smoke gates.

Quality Center and release gates

Link native Databricks and MLflow evidence, then combine it with BDD, live-suite, JUnit, and CLI results before production promotion.

Databricks-native AI quality

Keep MLflow datasets, scorers, traces, feedback, artifacts, and lineage in Databricks while Fabric adds TypeScript automation and release policy.

Enterprise controls available today

Configure tenant roles, scoped API keys, production Quality policy, retention, legal hold, SSO/SCIM, and audit forwarding.

Self-host on any static CDN

fx publish --static writes a signed manifest, verification files, and tag-loader bundle to a directory. Upload to S3, R2, GCS — no manifest-worker required.

Analyst tools

Analyst-ready warehouse roll-ups: standalone SQL, a self-contained dbt project, and .Rmd notebooks for CR lift, Bayesian posteriors, and SRM χ².

What you get

• GitOps authoring — define experiments in YAML and apply them with the fx CLI.
• Hosted API — Next.js control plane with typed routes, org tenancy, audit events, manifests, and aggregate reporting.
• Studio UI — operational Overview, six-stage Pipeline, guided authoring, experiment analytics, Quality Center test evidence, audit logs, and enterprise settings.
• Better Auth — email/password, magic links, OAuth, SAML/OIDC SSO, SCIM, organizations, API keys, admin, and 2FA.
• Postgres database — standard Postgres schema; Neon recommended for hosted/serverless staging.
• Cloudflare edge — manifest delivery from KV, exposure/conversion ingestion to R2, scheduled audit export worker, and this docs site.
• Databricks validation — TypeScript BDD, live target checks, restricted identities, isolated resources, cleanup, and schema-versioned Quality evidence.
• Warehouse aggregation — Databricks production path plus local warehouse adapter for tests.
• Agents — SRM watcher and analyst workflows on Fabric Harness + Temporal.

High-level architecture

Current status

Production deployment active — the combined Studio and hosted API run on Databricks Apps over standard Postgres (Lakebase in the Fabric-managed deployment; the control plane is Postgres-provider-agnostic — see Database). Runtime delivery, signed previews, migration tooling, Studio analytics, Databricks BDD, live artifact validation, Quality Center, and self-hosted publishing are available.

Core Azure data-engineering targets are live-gated. Available packs and their exact limitations are listed in Databricks target packs. Compatibility is documented capability by capability—including intentional differences and uncertified targets—in the compatibility matrix.