Every Microsoft Fabric Feature, Explained

Hot, cool, and cold tiers matching Azure Storage, plus lifecycle rules that move files by creation, modification, or last-access time, set per workspace default or per file.

Role-based access control defined once on the data and enforced consistently across Spark, the SQL analytics endpoint, and Direct Lake over OneLake, with the union of roles resolved for each user; the Members and Data views are GA September 2026.

Turns structured files behind a shortcut into managed Delta tables that Fabric Spark keeps in sync by polling every two minutes, with no pipeline to build; Excel input is still in preview.

A Windows client that mounts every workspace you can access as a folder tree, syncing placeholders on demand the way OneDrive does, with a tenant setting to switch it off.

Metadata virtualization that generates a virtual Delta log for Iceberg tables and virtual Iceberg metadata for Delta tables, so both formats are readable by every Fabric engine and by external Iceberg readers.

Workspace-level streaming of data-access events as JSON into a lakehouse you choose, with optional immutable retention for compliance and a documented schema you can turn into Delta tables.

The tenant-wide Explore, Govern, and Secure surface for Fabric items (the Govern view is GA September 2026), also embedded in Teams, Excel, and Copilot Studio, with a search REST API for programmatic discovery.

Cross-tenant, in-place sharing that creates a read-only shortcut in the consumer's lakehouse pointing back at the source, so changes show immediately and revoking the share cuts access.

The single logical data lake behind every Fabric workload, built on ADLS Gen2 and storing tables as Delta or Iceberg, so one copy of data serves Spark, T-SQL, KQL, and Power BI without duplication.

Pointers that make external or cross-workspace data appear as local folders and tables, with schema kept in sync automatically and an optional cache for cross-cloud reads.

A managed dbt Core runtime that builds, tests, and documents SQL models against Fabric Data Warehouse and other supported targets, enabled per tenant from the admin portal, where the switch is still labelled "dbt jobs (preview)".

The Mapping Data Flow authoring canvas and Spark execution engine embedded in Dataflow Gen2, with a built-in migration experience from Azure Data Factory and Synapse.

Natural-language authoring for Dataflow Gen2 and pipelines, including pipeline generation, expression building, and error explanation, gated by a tenant setting and a paid F2 or P1 capacity.

Mounts an existing factory as a native Fabric item for side-by-side operation, paired with an assessment-first upgrade experience (in preview since March 2026) that grades each pipeline Ready, Needs review, Coming soon, or Not compatible; eligible pipelines that contain Mapping Data Flows now upgrade to Mapping Data Flow transforms in Dataflow Gen2, which are themselves preview.

A hosted Airflow (2.10.5, Python 3.12) with Git sync, autoscale, high availability, and deferrable operators, running as a Fabric item alongside pipelines.

A guided, serverless data-movement item with watermark and CDC incremental modes, append, overwrite, merge, and SCD Type 2 writes, audit columns, and multiple schedules; CDC replication and SCD Type 2 are GA September 2026, while auto-partitioning and the Eventstream source and destination are still in preview. Learn's Copy job page still labels CDC replication preview.

The low-code orchestration engine, carrying the Copy activity plus notebook, Spark, stored procedure, dataflow, Teams, and email activities, with the same SLA as Azure Data Factory; retry backoff and the Lakehouse maintenance activity (VACUUM and OPTIMIZE as a pipeline step) are GA September 2026.

Low-code ETL built on Power Query that loads to lakehouse, warehouse, KQL, SQL database, ADLS Gen2, Snowflake, PostgreSQL, Excel workbooks, and more; since April 2026 every new Gen2 dataflow is created with CI/CD and Git support. Learn now calls Power BI's Dataflow Gen1 legacy: still supported, with no new feature investment.

The connector library shared by pipelines, Copy job, and Dataflow Gen2, reached through direct cloud connections, the on-premises data gateway, or the virtual network data gateway; tenant-wide authentication, tenant allowlists, and connection Admin APIs are GA September 2026.

An autonomous agent that runs persistent, Fabric-hosted tasks against a lakehouse you choose. It creates and tests Spark notebooks and Delta tables under settings you review first, and anyone who can open that lakehouse can see its progress.

The current-generation Spark runtime, generally available but opt-in per workspace or environment, with the native execution engine included and Delta 4.x-only features marked experimental.

Declarative, dependency-aware Delta tables that Fabric refreshes incrementally when source tables have change data feed enabled, with constraints, lineage, and run analytics built in; PySpark authoring is still in preview.

Serverless Python 3.11 functions with PyPI packages, managed connections to Fabric data sources, and an auto-generated REST endpoint with Entra authentication.

An opt-in, per-capacity model (announced as Autoscale Billing for Spark) that moves Spark jobs onto separately billed serverless compute under a CU ceiling you set, with interactive jobs throttled and batch jobs queued at the limit.

A vectorized C++ engine built on Velox and Gluten that offloads supported operators from JVM Spark with no code changes, enabled per environment, notebook, or job, and falling back automatically when an operator is unsupported.

A workspace item bundling runtime version, compute sizing, libraries (Quick or Full publish mode), and resource files that notebooks and Spark job definitions attach to, shareable across workspaces on the same capacity.

A generated GraphQL schema, resolvers, and endpoint over Fabric data sources, with relationships, stored procedures, an in-portal editor, and CI/CD support (GA February 2026).

Delta tables plus files in one item, with automatic table discovery, a read-only SQL analytics endpoint generated for free, and schema support (GA since December 2025) for organizing tables.

The primary code item for engineering and data science, with public APIs for CI/CD (GA March 2026), Python-only notebooks that skip Spark entirely (GA September 2025), and a security context that follows whoever last edited the pipeline or schedule.

A code item that submits a main definition file plus libraries and arguments to the Spark cluster against a default lakehouse, runnable from pipelines with a service principal or workspace identity.

Starter pools give a session in 5 to 10 seconds from prewarmed medium nodes on a best-effort basis; custom pools set node size (4 to 64 vCores per node), autoscale bounds, dynamic executor allocation, and session expiry, sized against two Spark vCores per capacity unit with a 3x burst; custom live pools (GA September 2026) keep a hydrated cluster warm on a schedule for roughly five second starts.

A guided experience that migrates tables, views, procedures, and security objects from a DACPAC, a zip of .sql or BTEQ files, or a direct source connection (direct connection GA June 2026), prioritizes failed scripts by dependency, and hands data movement to Copy job.

System restore points every eight hours plus user-defined ones with configurable retention (1 to 120 days), read-only snapshots you can roll forward (GA November 2025), and dropped-warehouse recovery from the recycle bin (GA July 2026).

Schema-aware T-SQL generation and repair in a chat pane, completions, and ribbon quick actions, gated by the tenant Copilot setting and a paid capacity.

An enterprise-scale relational warehouse on Delta storage with separated compute and storage, cross-database queries, COPY INTO and CTAS loading, autonomous workload management, IDENTITY columns (GA August 2026), result set caching on by default (GA September 2026), and custom SQL pools, which Learn no longer labels preview; the whats-new preview table still lists them.

The auto-provisioned, read-only T-SQL surface over Delta tables, shared by lakehouses, SQL databases, and mirrored databases, where OneLake security rules are enforced for SQL clients.

The in-portal editor for warehouses, SQL analytics endpoints, and mirrored databases, with IntelliSense, templates, autosave, cross-warehouse queries, and a September 2026 GA wave adding a faster grid, bulk query management, and .sql import and export.

Named, schema-governed events (GA September 2026) published from Activator, Eventstream, notebooks, or user data functions into Real-Time hub, where Activator and Eventhouse consume them without touching the publisher; every published event is also written to a KQL table in an eventhouse by default for history and dashboards, and the Schema Registry in Real-Time hub holds the contract.

A workspace item that groups Avro schemas and event types on the xRegistry model, keeps numbered versions you can compare, and feeds a tenant-wide registry page in Real-Time hub that business events and eventstreams read their contracts from.

KQL-backed dashboards with live refresh, Copilot-generated tiles and queries, conditional formatting, richer KPI tiles with reference lines, text, and divider visuals (GA September 2026), Git integration, and safe sharing that does not expose the underlying data.

A first-class map item that layers Eventhouse KQL results and lakehouse spatial data with points, lines, polygons, and imagery, refreshing on an interval and billed as ordinary capacity consumption.

A low-code ontology and mapping experience for physical environments, stored in OneLake, with an explorer, time-series views, public APIs, and hooks into Power BI, Real-Time Dashboard, data agents, and Activator.

The automatically provisioned home for data in motion, with connectors for Event Hubs, IoT Hub, Service Bus, six database CDC sources, Kafka variants, Kinesis, Pub/Sub, and Fabric capacity, workspace, OneLake, and job events.

A no-code detection engine that evaluates stateless and stateful rules per business object, previews how often a rule would have fired, and only bills while rules are running; rules can also be authored inside Eventstream (GA August 2026); copy job, dataflow, and function actions and a visible KQL or SQL source query per rule are GA September 2026.

No-code stream collection and processing with a Kafka-compatible endpoint, a SQL operator (GA April 2026), and DeltaFlow (preview since March 2026), which turns raw CDC events into analytics-ready, schema-tracked streams. Learn's Eventstreams overview still labels the SQL operator preview.

A container of KQL databases that indexes data by arrival time for sub-second queries over billions of events, exposes one logical copy to OneLake, and now includes native anomaly detection (preview since September 2025) and a Capacity Planner that schedules minimum capacity by hour and day (no longer labelled preview as of October 2026), plus customer-managed keys and an entity diagram (GA September 2026).

FLAML-based automated model selection and hyperparameter tuning with Spark-parallel trials, constraint and metric configuration, Plotly visualizations, and a low-code experience that reached GA in March 2026.

The SemPy library and Spark native connector that read semantic models into a FabricDataFrame carrying measures and metadata, with validation tools for relationships and functional dependencies.

Nine LLM-powered functions callable from pandas, PySpark, T-SQL, and Dataflow Gen2 with high concurrency and row-level isolation, defaulting to a built-in model and configurable to Azure OpenAI or Foundry models; multimodal input is in preview.

A visual data-prep surface launched from any notebook cell, with a searchable operation library, PROSE-powered suggestions, Copilot prompts, and pandas-to-PySpark translation of the applied steps.

Built-in MLflow experiments and runs with visual comparison, tagging, MLflow 3 logged models and GenAI traces, cross-workspace logging (GA April 2026), and a Save as ML model path into the registry.

A free observability pane in the Fabric portal that discovers database resources through your existing Azure and Fabric permissions and shows issues, suggestions, and performance across them in one estate view.

A guided migration that imports schema from a DACPAC, flags and fixes objects SQL database in Fabric cannot run, copies data with a Copy job, and then re-enables constraints, triggers, and indexes.

The Azure SQL Database engine as a SaaS Fabric item, with automatic tuning, near-real-time replication to OneLake, Git integration, SqlPackage portability, a GraphQL API in one click, DiskANN vector indexes (GA September 2026), and its own Copilot (GA November 2025).

The Azure Cosmos DB for NoSQL engine with autonomous defaults, Entra-only identity, and automatic Delta mirroring into OneLake so notebooks, Power BI, and cross-database queries see the same documents.

Mirrors only the catalog structure and reaches the data through shortcuts, with an auto-generated SQL analytics endpoint and Direct Lake support; Dremio catalogs, AWS Glue, and Google Lakehouse Runtime Catalog use the same pattern in preview.

A landing-zone contract in OneLake where any writer drops inserts, updates, and deletes per the open mirroring format, and the replication engine turns them into a mirrored database with a SQL analytics endpoint.

Managed, near-real-time replication of whole databases or selected tables into Delta tables with free replication compute and a storage allowance tied to capacity size; SQL Server 2025, PostgreSQL, and Cosmos DB reached GA in November 2025, Oracle and SAP in March 2026, BigQuery in August 2026, and SharePoint lists, change data feed, and view mirroring in September 2026. Learn's mirroring overview still labels SharePoint lists preview.

Save as PBIP to get folder-per-item text definitions with public JSON schemas; the project format and the enhanced report format (PBIR, the default since January 2026) are generally available, as is TMDL view in Desktop (GA September 2025), while live reload of external edits is still in preview.

Apps as first-class Fabric items, multiple per workspace, with audiences as child items, automatic access propagation to included items and their semantic models, and Teams and PowerPoint support.

A storage mode that pages Delta columns straight into VertiPaq memory with metadata-only refreshes, reading any Fabric Delta source without DirectQuery fallback; calculated columns and tables on it are still in preview.

Data function buttons that pass the current filter context to a user data function, which writes back or triggers downstream systems, with in-report alerts built on the same pattern.

The report Copilot pane (generally available) plus preview surfaces for standalone chat across all your items, app-scoped answers with author-verified responses, and model-authoring help, all billed as capacity consumption.

The tabular model at the center of Fabric BI, created explicitly since September 2025 (default models are no longer auto-generated) and editable on the web (GA September 2025), with Direct Lake as the storage mode for new models on lakehouses and warehouses.

Direct Lake over a single lakehouse or warehouse: reads its Delta tables, uses its SQL analytics endpoint for table and view discovery and permission checks, and falls back to DirectQuery when it cannot read the Delta files directly (SQL views, SQL-based granular access control). The model's Direct Lake behavior property decides whether that fallback happens or the query fails.

Power BI data answering inside Microsoft 365 Copilot Chat: Copilot finds a report or semantic model the user can already open, queries it with their permissions, and reconciles the answer with their files, chats, and email.

An enterprise planning suite inside Fabric IQ with planning sheets, PowerTable sheets for dimensional models, Intelligence sheets for variance analysis, and Infobridge for integration, writing results back to a SQL database in Fabric.

A scale-out labeled property graph over OneLake tables with ISO GQL and natural-language-to-GQL querying, built-in algorithms, auto-shutdown when idle, and data agent support for graph reasoning (preview).

Playbook-driven agents with their own Microsoft Entra Agent ID, running delegated as their creator, evaluating state and transition rules every five minutes, and drafting rules conversationally through Copilot.

Three tiers of Model Context Protocol server: the open-source local server for API specs, item schemas, and OneLake files, the hosted Core server (GA September 2026) for workspaces, items, and permissions, and the hosted Fabric IQ server (GA September 2026) for read-only DAX over Power BI semantic models; the per-workload remote servers for Data Warehouse, Eventhouse and Activator, operations agents, and ontology are still in preview.

Configurable conversational analytics that translates questions to SQL, DAX, KQL, or Graph queries under the caller's permissions and Purview policies, publishable to Microsoft 365 Copilot (GA June 2026), Copilot Studio (GA August 2026), and Teams, with Git and deployment-pipeline support, plus chart answers and long-running MCP tasks (both GA September 2026). The preview integration with Microsoft Foundry is being retired: from 6 November 2026 it is no longer supported, and Foundry agents that use it must move to the Fabric IQ (OneLake Catalog) tool before then.

Microsoft's enterprise intelligence layer for business data, sitting alongside Work IQ and Foundry IQ, with three layers: unified data in OneLake, curated KPIs in semantic models, and operational context in ontologies; its first GA surfaces arrived in September 2026 (Power BI answers in Microsoft 365 Copilot Chat and the Fabric IQ MCP server).

Entity types, properties, relationships, natural-language rules, DAX measures from Power BI, namespaces, inheritance, versions, and RDF and OWL import, with data bindings queried live at the source (no instance graph to refresh) and a built-in ontology agent for natural-language querying.

Per-workload assistants on a shared Azure OpenAI architecture, on by default for paid capacities, billed as background capacity consumption, and gated by cross-geo tenant settings outside the US and EU data boundary.

Allow-only rules stored as a workspace item and activated per tenant or capacity scope, managed from the Policies center in the OneLake catalog, REST APIs, and Git; three policies today cover item creation, workspace settings edits, and external data sharing, and every change is recorded in the audit log.

Soft delete on by default with a three-day retention, configurable to 90 days by tenant admins, alongside collaborative-workspace retention of 7 to 90 days.

Tenant-level and workspace-level private links (workspace level GA October 2025), managed virtual networks and private endpoints for Data Engineering, trusted workspace access, conditional access, and service tags.

Deny-by-default egress per workspace with allow lists via managed private endpoints or data connection rules, covering Spark, Data Factory, Warehouse, Real-Time Intelligence, mirroring, Fabric IQ agents, and OneLake, plus a tenant admin API to audit every policy; protection for OneLake shortcuts and copy operations and a networking communication policies Admin API reached GA in September 2026, and as of October 2026 coverage for semantic models, Activator, and Eventhouse is GA too.

Workspace-level customer-managed keys that wrap the Microsoft encryption key, extended to SQL database in Fabric (GA March 2026) and manageable through APIs (GA August 2026).

Sensitivity labels with default and inherited labeling and public APIs (GA March 2026), Purview DLP policies for Fabric and Power BI (GA September 2025) with the restrict-access action for structured OneLake data (GA September 2026), and every activity captured in Purview Audit.

Logical grouping of workspaces into domains and subdomains with domain admins and contributors, default-domain rules by user or group, and a slice of tenant settings that can be overridden per domain.

An automatically managed service principal per workspace that unlocks trusted workspace access to ADLS Gen2 and shows up in the admin portal's Fabric identities tab (GA January 2026) with Purview and Entra audit trails.

Promoted and certified endorsement with delegated certifiers, admin-defined tags on items (GA April 2025) and workspaces (GA March 2026), and workspace lineage with impact analysis.

A capacity setting, on by default when you create an F SKU capacity, that pays for usage beyond the SKU at the point throttling would otherwise start, checked every five minutes against a rolling 24-hour CU-hour threshold you set, which has to fit inside your Fabric quota and is not a hard spending cap; when the threshold is reached, normal throttling resumes. Surge protection still applies, and billed overages appear in the Capacity Metrics app, Azure Cost Management, and Real-Time hub capacity events.

A workspace item that puts items in ordered deployment groups with pre-deploy and post-deploy actions, versioned in Git and attached per operation to Git integration, deployment pipelines, or the REST APIs, stopping at the first failed item or action.

A monitoring item per workspace that sends telemetry from supported items into a read-only Eventhouse, its own or a shared one in the same region, queryable with KQL and feeding dashboards, Activator alerts, and operations agent investigations; collection starts switched off.

A managed application platform built on the Rayfin SDK where decorated TypeScript classes generate schema, endpoints, and row-level authorization, with TypeScript server-side functions, a secret store, and typed connectors to lakehouses, warehouses, SQL databases, and semantic models, deployed with rayfin up and billed against the workspace capacity.

The open-source fab command line (pip-installable, Python 3.10 or later) with interactive, service principal, certificate, and managed identity login; v1.5 (GA March 2026) added one-command workspace deployment with fabric-cicd, Power BI scenario support, a pre-installed CLI in notebooks, and Python 3.13.

Frontend Entra token access, item definition storage, per-item OneLake folders with shortcuts, standard CRUD APIs, iframe relaxation, and self-service workload publishing (both GA March 2026).

A workspace item holding typed variables and item references with one active value set per stage, consumed by pipelines, shortcuts, notebooks, Dataflow Gen2, Copy job, user data functions, and Plan.

The admin-installed monitoring app with Health, Compute, Storage, timepoint, and Spark autoscale pages, health-state duration, heatmap, and contributor drill-down views (GA September 2026), plus the Chargeback app (GA May 2026), alongside surge protection for background operations (GA June 2025) with workspace-level surge protection in preview.

Folders are organizational units inside a workspace (ten nesting levels, workspace-inherited permissions, most item types creatable inside them; Dataflow Gen2 and streaming items still land at the root). Task flows are a canvas above the item list with predefined or custom tasks typed by stage (get, mirror, store, prepare, analyze, track, visualize, distribute, develop) and connectors that document relationships rather than move data. A Folder REST API (preview, Apr 2025) automates folder management.

Workspace-level sync with Azure DevOps, GitHub, and GitHub Enterprise Cloud (data residency GA June 2026), with compare and commit, branch workspaces, and selective branching (all GA September 2026), preserving folder structure and honoring workspace network security rules.

Azure-billed F2 through F8192 capacities measured in capacity units, per-second billing with optional reservations, and workspace types that decide which items a workspace can hold; P SKUs are being retired in favor of F.

A cross-workspace Activities page listing job runs with status, duration, history, and errors, filterable by status, item type, time window, submitter, and workspace; the refreshed hub (preview), which Learn's page now documents, adds Alerts, Applications, Capacities, and Agents pages.

Stage-to-stage cloning of supported items with item pairing, folder hierarchy preservation, and REST APIs, with a redesigned UI in preview and enhanced-metadata semantic models required since February 2026.

The REST API foundation for item CRUD, definitions (bulk export and import GA September 2026), capacity, access, and jobs, the Terraform provider for Fabric (GA March 2025) for environments, and the fabric-cicd Python library, which the docs call the most widely adopted deployment tool.