Optimizing Spark Performance with the Native Execution Engine (NEE) in Microsoft Fabric

Spark tuning often starts with the usual suspects (shuffle volume, skew, join strategy, caching)… but sometimes the biggest win is simply executing the same logical plan on a faster engine.

Microsoft Fabric’s Native Execution Engine (NEE) does exactly that: it keeps Spark’s APIs and control plane, but runs a large portion of Spark SQL / DataFrame execution on a vectorized C++ engine.

What NEE is (and why it’s fast)

NEE is a vectorized native engine that integrates into Fabric Spark and can accelerate many SQL/DataFrame operators without you rewriting your code.

  • You still write Spark SQL / DataFrames.
  • Spark still handles distributed execution and scheduling.
  • For supported operators, compute is offloaded to a native engine (reducing JVM overhead and using columnar/vectorized execution).

Fabric documentation calls out NEE as being based on Apache Gluten (the Spark-to-native glue layer) and Velox (the native execution library).

When NEE tends to help the most

NEE shines when your workload is:

  • SQL-heavy (joins, aggregates, projections, filters)
  • CPU-bound (compute dominates I/O)
  • Primarily on Parquet / Delta

You’ll see less benefit (or fallback) when you rely on features NEE doesn’t support yet.

How to enable NEE (3 practical options)

1) Environment-level toggle (recommended for teams)

In your Fabric Environment settings, go to Acceleration and enable the native execution engine, then Save + Publish.

Benefit: notebooks and Spark Job Definitions that use that environment inherit the setting automatically.

2) Enable for a single notebook / job via Spark config

In a notebook cell:

%%configure
{
  "conf": {
    "spark.native.enabled": "true"
  }
}

For Spark Job Definitions, add the same Spark property.

3) Disable/enable per-query when you hit unsupported features

If a specific query uses an unsupported operator/expression and you want to force JVM Spark for that query:

SET spark.native.enabled=FALSE;
-- run the query
SET spark.native.enabled=TRUE;

How to confirm NEE is actually being used

Two low-friction checks:

  1. Spark UI / History Server: look for plan nodes ending with Transformer or nodes like *NativeFileScan / VeloxColumnarToRowExec.
  2. df.explain(): the same Transformer / NativeFileScan / Velox… hints should appear in the plan.

Fabric also exposes a dedicated view (“Gluten SQL / DataFrame”) to help spot which queries ran on the native engine vs. fell back.

Fallback is a feature (but you should know the common triggers)

NEE includes an automatic fallback mechanism: if the plan contains unsupported features, Spark will run that portion on the JVM engine.

A few notable limitations called out in Fabric documentation:

  • UDFs aren’t supported (fallback)
  • Structured streaming isn’t supported (fallback)
  • File formats like CSV/JSON/XML aren’t accelerated
  • ANSI mode isn’t supported

There are also some behavioral differences worth remembering (rounding/casting edge cases) if you have strict numeric expectations.

A pragmatic “NEE-first” optimization workflow

  1. Turn NEE on for the environment (or your job) and rerun the workload.
  2. If it’s still slow, open the plan and answer: is the slow part running on the native engine, or did it fall back?
  3. If it fell back, make the smallest possible change to keep the query on the native path (e.g., avoid UDFs; prefer built-in expressions; standardize on Parquet/Delta).
  4. Once the plan stays mostly native, go back to classic Spark tuning: reduce shuffle volume, fix skew, sane partitioning, and confirm broadcast joins.

References

This post was written with help from ChatGPT 5.2

The Best Thing That Ever Happened to Your Spark Pipeline Is a SQL Database

Here’s a counterintuitive claim: the most important announcement for Fabric Spark teams in early 2026 has nothing to do with Spark.

It’s a SQL database.

Specifically, it’s the rapid adoption of SQL database in Microsoft Fabric—a fully managed, SaaS-native transactional database that went GA in November 2025 and has been quietly reshaping how production data flows into lakehouse architectures ever since. If you’re a data engineer running Spark workloads in Fabric, this changes more than you think.

The ETL Pipeline You Can Delete

Most Spark data engineers have a familiar pain point: getting operational data from transactional systems into the lakehouse. You build ingestion pipelines. You schedule nightly batch loads. You wrestle with CDC (change data capture) configurations, watermark columns, and retry logic. You maintain all of it, forever.

SQL database in Fabric eliminates that entire layer.

When data lands in a Fabric SQL database, it’s automatically replicated to OneLake as Delta tables in near real-time. No pipelines. No Spark ingestion jobs. No orchestration. The data just appears, already in the open Delta format your notebooks and Spark jobs expect.

This isn’t a minor convenience—it’s an architectural shift. Every ingestion pipeline you don’t write is a pipeline you don’t debug at 2 AM.

What This Actually Looks Like in Practice

Let’s say you’re building an analytics layer on top of an operational SaaS application. Today, your architecture probably looks something like this:

  1. Application writes to Azure SQL or Cosmos DB
  2. ADF or Spark job pulls data on a schedule
  3. Data lands in a lakehouse as Delta tables
  4. Downstream Spark jobs transform and aggregate

With SQL database in Fabric, steps 2 and 3 vanish. Your application writes directly to the Fabric SQL database, and the mirrored Delta tables are immediately available for Spark processing. Here’s what your downstream notebook looks like now:

# Read operational data directly — no ingestion pipeline needed
# The SQL database auto-mirrors to OneLake as Delta tables
orders_df = spark.read.format("delta").load(
    "abfss://your-workspace@onelake.dfs.fabric.microsoft.com/your-sqldb.SQLDatabase/dbo.Orders"
)

# Your transformation logic stays the same
from pyspark.sql import functions as F

daily_revenue = (
    orders_df
    .filter(F.col("order_date") >= F.date_sub(F.current_date(), 7))
    .groupBy("product_category")
    .agg(
        F.sum("total_amount").alias("revenue"),
        F.countDistinct("customer_id").alias("unique_customers")
    )
    .orderBy(F.desc("revenue"))
)

daily_revenue.write.format("delta").mode("overwrite").saveAsTable("gold.weekly_revenue_by_category")

The Spark code doesn’t change. What changes is everything upstream of it.

The Migration Risk Nobody’s Talking About

Here’s where it gets interesting—and where Malcolm Gladwell would lean forward in his chair. The biggest risk of SQL database in Fabric isn’t technical. It’s organizational.

Teams that have invested heavily in ingestion infrastructure will face a classic innovator’s dilemma: the new path is simpler, but the old path already works. The temptation is to keep running your existing ADF pipelines alongside the new auto-mirroring capability, creating a hybrid architecture that’s worse than either approach alone.

My recommendation: don’t hybrid. Pick a workload, migrate it end-to-end, and measure. Here’s a concrete rollout checklist:

  1. Identify a candidate workload — Look for Spark jobs whose primary purpose is pulling data from a SQL source into Delta tables. These are your highest-value migration targets.
  2. Provision a Fabric SQL database — It takes seconds. You provide a name; Fabric handles the rest. Autoscaling and auto-pause are built in.
  3. Redirect your application writes — Point your operational application to the new Fabric SQL database. The engine is the same SQL Database Engine as Azure SQL, so T-SQL compatibility is high.
  4. Validate the Delta mirror — Confirm that your data is appearing in OneLake. Check schema fidelity, latency, and row counts:
# In your Spark notebook, validate the mirrored data
spark.sql("""
    SELECT COUNT(*) as row_count,
           MAX(modified_date) as latest_record,
           MIN(modified_date) as earliest_record
    FROM your_sqldb.dbo.Orders
""").show()
  1. Decommission the ingestion pipeline — Once validated, turn off the ADF or Spark ingestion job. Don’t just disable it—delete it. Zombie pipelines are how technical debt accumulates.
  2. Update your monitoring — Your existing data quality checks should still work since the Delta tables have the same schema. But update your alerting to watch for mirror latency instead of pipeline run failures.

The AI Angle Matters for Spark Teams Too

There’s a second dimension to this announcement that Spark engineers should pay attention to: the native vector data type in SQL database supports semantic search and RAG patterns directly in the transactional layer.

Why does that matter for Spark teams? Because it means your embedding pipelines can write vectors back to the same database your application reads from—closing the loop between batch ML processing in Spark and real-time serving in SQL. Instead of maintaining a separate vector store (Pinecone, Qdrant, etc.), you use the same SQL database that’s already mirrored into your lakehouse.

This is the kind of architectural simplification that compounds over time. Fewer systems means fewer failure modes, fewer credentials to manage, and fewer things to explain to your successor.

The Rollout Checklist

  • This week: Inventory your existing ingestion pipelines. How many just move data from SQL sources to Delta?
  • This sprint: Provision a Fabric SQL database and test the auto-mirror with a non-critical workload.
  • This quarter: Migrate your highest-volume ingestion pipeline and measure CU savings.
  • Track: Mirror latency, CU consumption before/after, and pipeline maintenance hours eliminated.

SQL database in Fabric went GA in November 2025 with enterprise features including row-level security, customer-managed keys, and private endpoints. For the full list of GA capabilities, see the official announcement. To understand how this fits into the broader Microsoft database + Fabric integration strategy, read Microsoft Databases and Microsoft Fabric: Your unified and AI-powered data estate. For Spark-specific Delta Lake concepts, the Delta Lake documentation remains the authoritative reference.

The best thing about this announcement isn’t any single feature. It’s that it makes your Spark architecture simpler by removing the parts that shouldn’t have been there in the first place.

This post was written with help from Claude Opus 4.6

Lakehouse Table Optimization: VACUUM, OPTIMIZE, and Z-ORDER

If your Lakehouse tables are getting slower (or more expensive) over time, it’s often not “Spark is slow.” It’s usually table layout drift: too many small files, suboptimal clustering, and old files piling up.

In Fabric Lakehouse, the three table-maintenance levers you’ll reach for most are:

  • OPTIMIZE: compacts many small files into fewer, larger files (and can apply clustering)
  • Z-ORDER: co-locates related values to improve data skipping for common filters
  • VACUUM: deletes old files that are no longer referenced by the Delta transaction log (after a retention window)

Practical note: in Fabric, run these as Spark SQL in a notebook or Spark job definition (or use the Lakehouse maintenance UI). Don’t try to run them in the SQL Analytics Endpoint.

1) Start with the symptom: “small files” vs “bad clustering”

Before you reach for maintenance, quickly sanity-check what you’re fighting:

  • Many small files → queries spend time opening/reading lots of tiny Parquet files.
  • Poor clustering for your most common predicates (date, tenantId, customerId, region, etc.) → queries scan more data than they need.
  • Heavy UPDATE/DELETE/MERGE patterns → lots of new files + tombstones + time travel files.

If you only have small files, OPTIMIZE is usually your first win.

2) OPTIMIZE: bin-packing for fewer, bigger files

Basic compaction

OPTIMIZE my_table;

Target a subset (example: recent partitions)

OPTIMIZE my_table WHERE date >= date_sub(current_date(), 7);

A useful mental model: OPTIMIZE is rewriting file layout (not changing table results). It’s maintenance, not transformation.

3) Z-ORDER: make your filters cheaper

Z-Ordering is for the case where you frequently query:

  • WHERE tenantId = ...
  • WHERE customerId = ...
  • WHERE deviceId = ... AND eventTime BETWEEN ...

Example:

OPTIMIZE my_table ZORDER BY (tenantId, eventDate);

Pick 1–3 columns that dominate your interactive workloads. If you try to z-order on everything, you’ll mostly burn compute for little benefit.

4) VACUUM: clean up old, unreferenced files (carefully)

VACUUM is about storage hygiene. Delta keeps old files around to support time travel and concurrent readers. VACUUM deletes files that are no longer referenced and older than the configured retention threshold.

VACUUM my_table;

Two practical rules:

  1. Don’t VACUUM aggressively unless you understand the impact on time travel / rollback.
  2. Treat the retention window as a governance decision (what rollback window do you want?) not just a cost optimization.

5) Fabric-specific gotchas (the ones that actually bite)

Where you can run these commands

These are Spark SQL maintenance commands. In Fabric, that means notebooks / Spark job definitions (or the Lakehouse maintenance UI), not the SQL Analytics Endpoint.

V-Order and OPTIMIZE

Fabric also has V-Order, which is a Parquet layout optimization aimed at faster reads across Fabric engines. If you’re primarily optimizing for downstream read performance (Power BI/SQL/Spark), it’s worth understanding whether V-Order is enabled for your workspace and table writes.

A lightweight maintenance pattern that scales

  • Nightly/weekly: OPTIMIZE high-value tables (or recent partitions)
  • Weekly/monthly: Z-ORDER tables with stable query patterns
  • Monthly: VACUUM tables where your org’s time travel policy is clear

Treat it like index maintenance: regular, boring, measurable.

References

This post was written with help from ChatGPT 5.2

OneLake Shortcuts + Spark: Practical Patterns for a Single Virtual Lakehouse

If you’ve adopted Microsoft Fabric, there’s a good chance you’re trying to reduce the number of ‘copies’ of data that exist just so different teams and engines can access it.

OneLake shortcuts are one of the core primitives Fabric provides to unify data across domains, clouds, and accounts by making OneLake a single virtual data lake namespace.

For Spark users specifically, the big win is that shortcuts appear as folders in OneLake—so Spark can read them like any other folder—and Delta-format shortcuts in the Lakehouse Tables area can be surfaced as tables.

What a OneLake shortcut is (and isn’t)

A shortcut is an object in OneLake that points to another storage location (internal or external to OneLake).

Shortcuts appear as folders and behave like symbolic links: deleting a shortcut doesn’t delete the target, but moving/renaming/deleting the target can break the shortcut.

From an engineering standpoint, that means you should treat shortcuts as a namespace mapping layer—not as a durability mechanism.

Where you can create shortcuts: Lakehouse Tables vs Files

In a Lakehouse, you create shortcuts either under the top-level Tables folder or anywhere under the Files folder.

Tables has constraints: OneLake doesn’t support shortcuts in subdirectories of the Tables folder, and shortcuts in Tables are typically meant for targets that conform to the Delta table format.

Files is flexible: there are no restrictions on where you can create shortcuts in the Files hierarchy, and table discovery does not happen there.

If a shortcut in the Tables area points to Delta-format data, the lakehouse can synchronize metadata and recognize the folder as a table.

One documented gotcha: the Delta format doesn’t support table names with space characters, and OneLake won’t recognize any shortcut containing a space in the name as a Delta table.

How Spark reads from shortcuts

In notebooks and Spark jobs, shortcuts appear as folders in OneLake, and Spark can read them like any other folder.

For table-shaped data, Fabric automatically recognizes shortcuts in the Tables section of the lakehouse that have Delta/Parquet data as tables—so you can reference them directly from Spark.

Microsoft Learn also notes you can use relative file paths to read data directly from shortcuts, and Delta shortcuts in Tables can be read via Spark SQL syntax.

Practical patterns (what I recommend in real projects)

Pattern 1: Use Tables shortcuts for shared Delta tables you want to show up consistently across Fabric engines (Spark + SQL + Direct Lake scenarios via semantic models reading from shortcuts).

Pattern 2: Use Files shortcuts when you need arbitrary formats or hierarchical layouts (CSV/JSON/images, nested partitions, etc.) and you’re fine treating it as file access.

Pattern 3: Prefer shortcuts over copying/staging when your primary goal is to eliminate edge copies and reduce latency from data duplication workflows.

Pattern 4: When you’re operationalizing Spark notebooks, make the access path explicit and stable by using the shortcut path (the place it appears) rather than hard-coding a target path that might change.

Operational gotchas and guardrails

Because moving/renaming/deleting a target path can break a shortcut, add lightweight monitoring for “broken shortcut” failures in your pipelines (and treat them like dependency failures).

For debugging, the lakehouse UI can show the ABFS path or URL for a shortcut in its Properties pane, which you can copy for inspection or troubleshooting.

Outside of Fabric, services can access OneLake through the OneLake API, which supports a subset of ADLS Gen2 and Blob storage APIs.

Summary

Shortcuts give Spark a clean way to treat OneLake like a unified namespace: read shortcuts as folders, surface Delta/Parquet data in Tables as tables, and keep your project’s logical paths stable even when physical storage locations vary.

References

This post was written with help from ChatGPT 5.2