Every Fabric customer wants the same thing from Spark: faster jobs, lower latency, and predictable performance. But Spark in Microsoft Fabric doesn’t behave like Spark on other platforms. It’s not about tuning clusters—it's about understanding the Fabric-native execution model, leveraging platform features, and designing with intent.

In this session, Estera Kot and Santhosh Kumar Ravindran will reveal a battle-tested, platform-aware techniques and cases to make your Fabric Spark workloads faster, cheaper, and more consistent—based on real-life enterprise scenarios and best practices.

This session is ideal for data engineers, architects, and solution owners who are already working in Fabric and want to unlock tangible performance gains across their pipelines, notebooks, and Lakehouse processing jobs.

You’ll learn how to:

🔧 Diagnose Slow Spark Jobs in Fabric
Use Monitoring Hub, Spark UI, and Profiler to identify slow stages, skewed joins, excessive shuffling, memory overflows, and CU spikes.

🚀 Accelerate Ingestion and Processing
Optimize read/write paths to OneLake by leveraging V-Order clustering, and efficient Delta Lake file sizing. Avoid small file issues and IO bottlenecks.

⚙️ Exploit Fabric Optimizations
Tap into Native Execution Engine (NEE) and latest perf optimizations to offload common transformations and avoid JVM overhead. Learn which workloads benefit the most—and when fallback paths are hurting performance.

📦 Make the Most of Fabric Lakehouse
Architect Lakehouse tables to support fast upserts, deletes, and merges using Z-Order, Delta log optimization, and OneLake shortcut patterns.

💡 Design Jobs with the Fabric Runtime in Mind
Plan job structure, transformations, and output strategy based on CU consumption patterns, concurrency behavior, and Fabric’s auto-scaling logic.

🧠 Use Copilot for Performance Suggestions
Learn how Fabric Copilot surfaces performance insights, suggests optimization patterns, and flags bottlenecks directly in your notebooks and pipelines.

🔥 Avoid Top Spark Performance Killers in Fabric
Including: unbounded joins, broadcast overload, poor partitioning, over-parallelism, and inefficient notebook orchestration.