FabricFabricExperiments
Testing on Databricks

Databricks workload support matrix

Exact Fabric Experiments certification coverage, release gates, preview boundaries, and remaining live evidence.

“Supports Databricks” is too broad to be a useful release claim. Fabric Experiments publishes an exact capability matrix instead. A workload is release-ready only when its package contract, offline tests, fx targets diagnostics, required live check, and retained Quality evidence are all green for the named environment.

Workload matrix

Databricks workloadTarget packPackage statusAzure live evidenceRelease meaning
SQL warehouses and Delta Lakesql-deltaLive-gatedGreenStatement execution, contracts, recovery, performance, backup, and rollback
Lakeflow pipelines, Auto Loader, and Volumeslakeflow-ingestionLive-gatedGreenRefresh and governed ingestion behavior
Jobs, notebooks, dbt, Python/JAR/Spark/SQL tasksjobs-code, lakeflow-jobsLive-gatedGreenExisting and disposable multi-task orchestration
Streaming tables, CDC, and Lakeflow Connectstreaming-cdc-connectLive-gatedGreenFreshness, CDC correctness, and managed replication
Unity Catalog, secrets, and storageunity-storageLive-gatedGreenPositive and negative access controls plus Volume/secret behavior
Databricks Apps and Lakebaseapps-operationalLive-gatedGreenApp health/resources and operational Postgres behavior
AI/BI dashboards and Genieaibi-genieLive-gatedGreenPublished dashboard and authenticated conversation
MLflow experiments and UC model lifecyclemlflow-lifecycleLive-gatedGreenExperiment and exact registered model version
MLflow 3 tracing, agent evaluation, scorers, monitoringmlflow3-agent-qualityShipped previewWorkspace-gatedTwo-task native MLflow Job plus four durable SQL evidence assertions; built-in Safety monitor feedback must be observed
Feature Engineering and online servingfeature-engineering-servingLive-gatedGreenOffline table, materialization, and freshness
Model Serving and AI Gatewaymodel-serving-gatewayLive-gatedGreenReadiness, inference, and gateway configuration
AI Search, Vector Search, RAG, and served agentsvector-rag-agentsLive-gatedGreenIndex readiness, retrieval, and agent inference
Managed MCP servers and UC MCP Servicesmanaged-mcp-agentsLive-gatedGreenStreamable HTTP list/call plus an explicit denied call
Customer Harness agent enrollmentagent-services-enrollmentShipped previewWorkspace-gatedSigned control-plane enrollment and UC Agent Services ACLs
Agent evaluation deliveryagent-evaluation-pipelineLive-gatedGreenAPI/outbox/Temporal/judge/SQL/Quality path
Data Quality Monitoring and App OTeldata-quality-observabilityShipped previewBaseline-gatedDQM freshness/completeness plus UC logs, spans, and metrics
Federation, Delta Sharing, and Clean Roomsfederation-sharing-cleanroomsLive-gatedGreenGoverned data access and collaboration metadata
Network, cost, policy, and regional recoverysecurity-cost-drLive-gatedGreenLeast privilege, cost guardrail, and secondary workspace workload
Metric views, materialization, SQL alerts, and performancesemantic-analyticsShippedAwaiting live evidenceNative Job plus four durable semantic and operational assertions
AI Functions and batch/streaming inferenceai-functions-batch-inferenceShippedAwaiting live evidenceOutput, quality, and lineage assertions
Knowledge Assistant, multi-agent, and Responses APIagent-orchestration-responsesShipped previewWorkspace-gatedGrounded native orchestration plus durable evidence
AI Runtime training and fine-tuningai-runtime-trainingShipped previewWorkspace-gatedTraining, evaluation, and exact UC model assertions
Lakebase Autoscaling Online Feature Storelakebase-autoscaling-featuresShipped previewWorkspace-gatedOnline publication, Feature Serving, and freshness
Declarative Automation Bundlesdeclarative-bundlesShippedAwaiting live evidenceCandidate provenance, execution, and rollback
Predictive Optimization, liquid clustering, and Iceberglakehouse-optimizationShippedAwaiting live evidenceThree fail-closed table assertions
Marketplace and OpenSharingmarketplace-open-sharingShippedAccount-gatedPrivate provider/recipient and governed asset evidence

“Shipped” means the typed pack, fail-closed doctor/run behavior, mocks, tests, and documentation exist. It is intentionally not the same as “live-gated.” Fabric Experiments must not describe a shipped workload as live-gated until its required live checks produce retained evidence on the exact candidate. Preview packs may ship with explicit workspace-gated status; they are not part of the GA live-gated claim. Every advanced pack also requires the restricted-principal check; evidence produced with a developer token cannot satisfy the OAuth M2M scope.

Languages and custom workloads

The framework can launch and inspect Python, R, Scala, SQL, notebook, wheel, JAR, Spark-submit, dbt, and pipeline workloads through Databricks Jobs. The test author owns the workload-specific assertion; the target pack owns submission, terminal-state handling, evidence, and failure semantics. Language support does not mean Fabric reimplements each language runtime.

Agent responsibility boundary

Fabric Harness owns agent definitions, policies, tools, durable execution, and Databricks deployment integrations. Fabric Experiments owns enrollment, version admission, evaluation thresholds, certification, and Quality evidence.

Customer agents use defineAgent() and run in a customer-owned worker. The official Experiments worker never downloads arbitrary customer modules. An admin first pins an Ed25519 public key, then submits an immutable signed manifest. The customer worker reconciles the requested external Agent Service and grants through @fabric-harness/databricks; the Experiments API does not contain a general Databricks runtime SDK.

See the Fabric Harness Databricks guide and Your First Agent for agent authoring.

Preview and cloud boundaries

  • Managed MCP, Agent Services, Data Quality Monitoring, and Apps telemetry are Databricks preview or Beta surfaces. They require explicit configuration and fail closed when absent; Fabric does not silently downgrade them.
  • Data Quality Monitoring enrollment is not a health pass. A release requires a recent latest result with both freshness and completeness Healthy; a newly enabled table in Training remains blocked until Databricks builds its baseline.
  • Agent Services currently certifies registration, discovery, metadata, and permissions. Databricks does not expose Agent Services runtime invocation in the Beta, so no runtime-invocation claim is made.
  • Azure Databricks is the current named live environment. AWS and Google Cloud package compatibility is not a live certification claim until the same required matrix runs there.
  • Private Link, customer-managed keys, unusual network egress, regional SKUs, and customer data/model quality remain customer-topology certifications.

Release checklist

Before publishing a candidate:

  1. run repository, boundary, type, lint, build, docs, OpenAPI, and package verification gates;
  2. run every required target-pack check against the named Azure workspace;
  3. run the API-to-outbox-to-Temporal-to-Model-Serving-to-SQL evaluation path;
  4. verify fh, fx, and Fiber against candidate artifacts rather than global workstation installs;
  5. publish evidence to Quality Center and retain the CI artifact; and
  6. publish only after every required row above is green.

See Databricks target packs for configuration, the Databricks enablement checklist for account and fixture prerequisites, and Certification evidence for the latest observed results.

On this page