Modernising Enterprise Analytics: Moving from Legacy Analysis Services to Microsoft Fabric Direct Lake

 

Introduction

Enterprise analytics architectures are undergoing a fundamental transformation. For years, cloud-hosted tabular models like Azure Analysis Services (AAS) provided the gold standard for high-speed reporting and governed DAX models. However, as data estates scale toward multi-terabyte footprints, traditional tabular platforms begin to encounter friction: redundant data copies, rigid refresh schedules, high infrastructure maintenance costs, and capacity scaling limits

Microsoft Fabric addresses these architectural constraints with Direct Lake mode. By enabling semantic models to read Delta Parquet files directly from lake storage with VertiPaq-level performance, Direct Lake eliminates the traditional trade-off between speed and data movement.

This article presents an anonymised, real-world architectural scenario demonstrating how enterprise teams can migrate legacy tabular semantic models to Microsoft Fabric Direct Lake.

Architectural Shift: Import, DirectQuery, and Direct Lake

Modernising a tabular semantic model requires understanding where Direct Lake fits compared to existing options:

• Import Mode: Loads entire datasets into memory via the VertiPaq engine. While queries execute quickly, the model demands extensive data duplication, long ETL refresh windows, and memory-bound scaling limits.
• DirectQuery: Leaves data in source storage, generating real-time queries on the fly. It ensures fresh data but frequently suffers from high latency and query concurrency bottlenecks.
• Direct Lake Mode: Blends the best of both approaches. The VertiPaq engine directly accesses delta storage in the lake without duplicating data or waiting for ETL data movement, providing fast query speeds on fresh datasets.

Implementation and things to concider

Before migrating tables directly into Fabric, evaluate source dataset sizes and identify which entities suit Direct Lake storage:

• Fact Table Partitioning and Scoping: Multi-hundred-million or billion-row tables should be evaluated against business retention needs (e.g., scoping to a rolling 3-year window).
• Materialized Lake Views: For high-volume transaction entities, pre-aggregating or filtering data via Materialized Views keeps the model performant while remaining within Fabric capacity thresholds.
• Dimension Handling: High-frequency, large dimension tables are kept in Direct Lake, while static or small reference lookups can be integrated via import or views depending on operational needs.

Query Performance Benchmarks

Benchmarking query performance across representative analytical workloads shows a clear improvement when switching from legacy tabular instances to Microsoft Fabric Direct Lake:

Benchmarekd using the same query in two models

Direct Lake Model Benchmark

AAS Model Benchmark

Workload Condition	Legacy Tabular Model	Fabric Direct Lake Mode	Observed Impact
Cold Cache Avg Total Duration	1,250.00 ms	1,101.80 ms	~12% Faster
Cold Cache Storage Engine (SE)	1,075.60 ms	986.00 ms	~8% Faster
Warm Cache Avg Total Duration	166.40 ms	122.20 ms	~26% Faster
Warm Cache Storage Engine (SE)	0.00 ms	0.00 ms	Instant In-Memory Retrieval

Beyond raw query duration, the true architectural advantage lies in operational efficiency. When considering the substantial processing time, compute overhead, and memory locking required by traditional tabular models during heavy ETL cycles, Direct Lake mode offers far superior data readiness and availability. By reading Delta tables in place without redundant staging pipelines, business users gain immediate access to analytical data the moment it lands in the lake.

Conclustion

Direct Lake delivers a 26% improvement in warm analytical query execution while eliminating the overhead of dedicated semantic model refreshes.

Transitioning from dedicated Analysis Services tiers to Microsoft Fabric Direct Lake allows organisations to consolidate analytics workloads into a unified architecture, cut infrastructure duplication, and provide business users with near-real-time insight generation.