Microsoft Fabric Spark: Native Execution Engine now generally available
The Fabric Spark Native Execution Engine (NEE) is now generally available (GA) as part of Fabric Runtime 1.3. This C++-based vectorized engine (built on Apache Gluten and Velox) runs Spark workloads directly on the lakehouse, requiring no code changes or new libraries. It supports Spark 3.5 APIs and both Parquet and Delta Lake formats, so your existing Spark queries simply run faster. In internal tests, the engine has delivered dramatic speedups – Microsoft’s benchmarks showed roughly 4× faster queries on a 1 TB TPC-DS workload versus vanilla Spark, and in our own Fabric GA trials we have seen up to 6× end-to-end performance gains on representative big-data jobs. The NEE is well-suited for a wide range of workloads (from batch ETL to interactive data science) because it processes data in columnar form and minimizes JVM overhead.
Enabling the Native Execution Engine
Getting started with the NEE in Fabric is straightforward. You can enable it in your Spark environment or session in several ways:
Environment (UI)
In the Fabric portal and create a new Environment item from the New item option and navigate to the Acceleration tab, and check Enable native execution engine. Once saved and published, all Spark Job Definitions and Notebooks that are using the environment inherit the setting.
Spark configuration
For a single notebook or job, set the Spark property in your session configuration. For example, in a notebook cell add:
%%configure
{ “conf”: { “spark.native.enabled”: “true” } }
If using a Spark job definition, include the same property in the job’s configuration. The change takes effect immediately without having to restart your Spark session.
– Workspace default environment: As part of workspace setup, you can attach an environment as the workspace default (via Workspace settings → Data Engineering/Science → Environment). By making that environment the default, all new Spark workloads in the workspace will automatically use the native engine without per-job configuration.
GA Performance Enhancements
In this GA release we have incorporated a series of optimizations and new features for the Native Execution Engine. Key improvements include:
– Native Delta write accleration
– Optimized Delta snapshot creation
– Deletion vectors support
– Expanded Delta operations support
Together, these enhancements plug feature gaps from preview and unlock further acceleration across common data engineering workload patterns.
Measured Performance Gains
With the Native Execution Engine enabled, users will see substantial end-to-end speedups. For example, in internal benchmarks on typical data aggregation and join queries, we have observed up to 6X faster runtimes compared to the standard Spark engine.
No Extra Cost for Big Gains
Importantly, the Fabric Native Execution Engine is included at no additional cost – just enable it and your existing Spark credit rates apply. Customers benefit from the dramatically faster execution without changing their spending plan: effectively you pay less for the same work.
Learn more about the latest performance updates as enabled as part of this General Availability release from our Native execution engine for Fabric Spark documentation.
