FabricFabric
Testing on Databricks

Certification evidence

Current public-package, Azure Databricks, browser data-path, and promotion evidence—with target boundaries and observation status.

This page separates three claims that are easy to confuse:

  • Shipped means the capability exists in a published artifact and has automated coverage.
  • Live-gated means Fabric has executed it against the named infrastructure and retained structured evidence.
  • Certified for promotion means the complete scheduled observation window has passed without a failure.

The exact capability status is maintained in the compatibility matrix. This page records the current operational evidence behind it.

Publicly reproducible evidence

EvidenceCurrent resultHow to reproduce
Published package installationCurrent npm releases pass package, live, and registry verificationInstall the versions under evaluation in a clean TypeScript project and run fx targets list; the release workflow also installs and executes the published CLI from an empty project.
External consumer CIPassed on July 17, 2026View the public workflow run.
DocumentationPublished with the current releaseOpen this page, the BDD quickstart, the Harness/Fiber guide, and the Quality Center guide; all are served from Cloudflare.

The consumer repository is deliberately separate from the product monorepo. It detects unpublished workspace dependencies, missing exports, peer-dependency errors, and tarballs that behave differently from source builds.

Fabric-operated live evidence

The summary below gives evaluators the target and boundary without presenting private CI artifacts as customer-accessible links.

GateLast verifiedResult and scope
Azure serverless required suiteJuly 15, 2026Passed SQL, Delta, Unity Catalog, Lakeflow, Volume, Lakebase, Databricks App, notebook, dbt, performance, failure-recovery, rotation, backup/restore, and rollback checks using a restricted workload identity.
Azure classic computeJuly 15, 2026Passed a disposable single-node, single-user notebook probe on DBR 16.4.x-scala2.12 with Standard_D2ds_v6; cleanup passed.
Browser-to-Studio pathJuly 15, 2026Passed browser emission → Cloudflare Queue/R2 → Auto Loader/Delta → aggregate API → authenticated Studio, including Bayesian, CUPED, time/depth, and segmented result assertions.
Clean-checkout Cloudflare staging deploymentJuly 15, 2026API, Studio, and edge workers built and deployed from a clean checkout.
Tenant-scoped Quality Center publishingJuly 16, 2026Passed BDD, Databricks live-suite, and generic CLI evidence publication using Databricks OAuth at the App gateway plus an organization-bound, expiring Fabric API key. The verification key was revoked afterward.
Fabric Harness operatorJuly 17, 2026Unit and integration gates pass tenant roles, dual-auth forwarding, loopback credential isolation, dry-run defaults, and audit redaction. Scheduled Cloudflare and Databricks Apps runs publish Fabric Harness operator evidence to Quality Center.
Lakeflow Jobs target packJuly 16, 2026Passed with a restricted service principal authenticated through Databricks OAuth M2M. The disposable Azure serverless matrix covered task values, explicit retry, condition branching, for_each, injected failure, targeted repair, cancellation, task outcomes, and fixture cleanup. JSON and JUnit evidence were produced.
Managed data and AI packsJuly 16, 2026A dedicated OAuth M2M service principal passed published AI/BI dashboard and Genie, MLflow model lifecycle, feature materialization and freshness, Model Serving and AI Gateway, Vector Search retrieval, and agent invocation checks.
Disposable advanced infrastructureJuly 16, 2026Streaming tables, AUTO CDC, query-based Lakeflow ingestion, federation, Delta Sharing, a two-metastore Clean Room, an enabled TEST-NET access list, compute policy, billing assertion, and a second-region serverless SQL assertion passed with cleanup. Direct PostgreSQL CDC is not included.
Customer-owned Harness worker AppJuly 20, 2026Deployment 01f18474078f1c8c8d66008156b34ec7, built from commit f9aaee4d91f1e76b9b63b3fda1d03482d934bca1, bootstrapped all three immutable generation-3 built-ins at @fabric-harness/sdk@4.2.4 and @fabric-harness/databricks@4.4.1. The three Temporal pollers and both outbox consumers became ready, authenticated /healthz and /readyz returned 200, and the first scheduled SRM run completed with scanned=1, errors=0. The replaced snapshot also drained all three workers cleanly on SIGTERM before the new deployment became ready. Recent logs, metrics, and spans continue exporting to dedicated Unity Catalog tables.
Managed MCP and agent evaluationJuly 20, 2026Managed MCP list/call/denial evidence passed. The agent-evaluation Job passed API submission, durable outbox dispatch, Temporal execution, Model Serving judge calls, SQL score persistence, and terminal Quality evidence.

Current Databricks promotion blockers

The following checks are implemented and intentionally fail closed in the named Azure workspace. They are not replaced with mocks or weaker assertions:

GateCurrent observationRequired resolution
MLflow 3 production monitoringJob 877870379677819, task run 370191088354931, registered and started the built-in Safety scorer. Its synchronous Safety preflight passed with databricks:/databricks-llama-4-maverick, and 19 post-activation traces were visible, but Databricks attached no asynchronous feedback within 2,400 seconds.Enable or repair the workspace's MLflow production-monitoring Beta/backend and obtain a real asynchronous feedback assessment.
Data Quality MonitoringEnrolled certification tables remain in Training; no latest table result has both freshness and completeness Healthy.Allow Databricks to finish the historical baseline, which can take up to two weeks, then rerun the fail-closed assertion.
Unity Catalog Agent ServicesList succeeds, but create returns AgentServices feature is not available.Have Databricks enable the Agent Services Beta for the workspace, then run signed customer enrollment and verify external-service metadata plus EXECUTE grants.
Genie Responses APIThe existing Genie conversation path is green; Agent Mode Responses returns FEATURE_DISABLED.Enable Genie Responses API for the workspace before adding that optional invocation claim.
Provisioned-throughput authoringNo dedicated provisioned-throughput endpoint is allocated to the certification environment.Approve the dedicated endpoint cost and fixture before requiring this optional SKU-specific check.

No package or documentation release may describe these rows as live-gated until the corresponding retained evidence is green. Workspace enablement, baseline training, and paid capacity are external prerequisites rather than conditions the Experiments code can bypass.

Evidence artifacts include JSON, JUnit XML, BDD HTML, rerun manifests, captured failure attachments, resource identifiers, durations, and cleanup results. Secrets are redacted before evidence is written.

The advanced suite runs weekly in two groups. Managed checks exercise named, reusable fixtures with the dedicated certification identity. Infrastructure checks create uniquely named resources, publish evidence to Quality Center, retain CI artifacts, and sweep those resources even after failure. This cadence detects permission drift, deleted fixtures, API changes, and cleanup failures; it does not replace customer acceptance testing on the customer's topology.

Observation window

The full staging path is currently certifying. Manual green runs prove the path works but do not increment the consecutive-night counter. The retained staging-certification.json workflow artifact is the source of truth for the live count. The Fabric-operated technical promotion gate remains blocked until seven scheduled runs pass in sequence; a scheduled failure resets the counter.

As of July 19, 2026, four consecutive scheduled runs (July 16–19) are green. The earliest possible completion is after the scheduled July 22 run, assuming the sequence remains green.

This distinction prevents a one-time successful demo from being presented as operational maturity.

Certified targets

Cloud and computeStatus
Azure Databricks serverless SQL and managed workloadsLive-gated
Azure classic job cluster: DBR 16.4 LTS, single-node Standard_D2ds_v6, single-userLive-gated
AWS DatabricksNot a certified target
Google Cloud DatabricksNot a certified target

Similar REST APIs do not establish cloud portability. AWS or Google Cloud moves to certified only after identity, storage, node types, networking and the complete live suite have evidence on that target.

What evaluators should verify themselves

Fabric's evidence establishes the release baseline. Before production adoption, run these checks in your own workspace:

  1. Complete the local and live quickstart.
  2. Add one representative Job, pipeline, notebook or dbt scenario.
  3. Prove an allowed and denied Unity Catalog access path.
  4. Retain JUnit, HTML and JSON evidence in your CI system.
  5. Run the required checks repeatedly under expected workload contention.
  6. Exercise backup/restore and rollback on disposable objects.

Use the advanced playbook to design those gates and the live-check reference for configuration.

On this page