🧭🏭Business Intelligence: Microsoft Fabric (Part VI: Data Stores Comparison)

Business Intelligence Series

Microsoft made available a reference guide for the data stores supported for Microsoft Fabric workloads [1], including the new Fabric SQL database (see previous post). Here's the consolidated table followed by a few aspects to consider: 

Area	Lakehouse	Warehouse	Eventhouse	Fabric SQL database	Power BI Datamart
Data volume	Unlimited	Unlimited	Unlimited	4 TB	Up to 100 GB
Type of data	Unstructured, semi-structured, structured	Structured, semi-structured (JSON)	Unstructured, semi-structured, structured	Structured, semi-structured, unstructured	Structured
Primary developer persona	Data engineer, data scientist	Data warehouse developer, data architect, data engineer, database developer	App developer, data scientist, data engineer	AI developer, App developer, database developer, DB admin	Data scientist, data analyst
Primary dev skill	Spark (Scala, PySpark, Spark SQL, R)	SQL	No code, KQL, SQL	SQL	No code, SQL
Data organized by	Folders and files, databases, and tables	Databases, schemas, and tables	Databases, schemas, and tables	Databases, schemas, tables	Database, tables, queries
Read operations	Spark, T-SQL	T-SQL, Spark*	KQL, T-SQL, Spark	T-SQL	Spark, T-SQL
Write operations	Spark (Scala, PySpark, Spark SQL, R)	T-SQL	KQL, Spark, connector ecosystem	T-SQL	Dataflows, T-SQL
Multi-table transactions	No	Yes	Yes, for multi-table ingestion	Yes, full ACID compliance	No
Primary development interface	Spark notebooks, Spark job definitions	SQL scripts	KQL Queryset, KQL Database	SQL scripts	Power BI
Security	RLS, CLS**, table level (T-SQL), none for Spark	Object level, RLS, CLS, DDL/DML, dynamic data masking	RLS	Object level, RLS, CLS, DDL/DML, dynamic data masking	Built-in RLS editor
Access data via shortcuts	Yes	Yes	Yes	Yes	No
Can be a source for shortcuts	Yes (files and tables)	Yes (tables)	Yes	Yes (tables)	No
Query across items	Yes	Yes	Yes	Yes	No
Advanced analytics	Interface for large-scale data processing, built-in data parallelism, and fault tolerance	Interface for large-scale data processing, built-in data parallelism, and fault tolerance	Time Series native elements, full geo-spatial and query capabilities	T-SQL analytical capabilities, data replicated to delta parquet in OneLake for analytics	Interface for data processing with automated performance tuning
Advanced formatting support	Tables defined using PARQUET, CSV, AVRO, JSON, and any Apache Hive compatible file format	Tables defined using PARQUET, CSV, AVRO, JSON, and any Apache Hive compatible file format	Full indexing for free text and semi-structured data like JSON	Table support for OLTP, JSON, vector, graph, XML, spatial, key-value	Tables defined using PARQUET, CSV, AVRO, JSON, and any Apache Hive compatible file format
Ingestion latency	Available instantly for querying	Available instantly for querying	Queued ingestion, streaming ingestion has a couple of seconds latency	Available instantly for querying	Available instantly for querying

It can be used as a map for what is needed to know for using each feature, respectively to identify how one can use the previous experience, and here I'm referring to the many SQL developers. One must consider also the capabilities and limitations of each storage repository.

However, what I'm missing is some references regarding the performance for data access, especially compared with on-premise workloads. Moreover, the devil hides in details, therefore one must test thoroughly before committing to any of the above choices. For the newest overview please check the referenced documentation!

For lakehouses, the hardest limitation is the lack of multi-table transactions, though that's understandable given its scope. However, probably the most important aspect is whether it can scale with the volume of reads/writes as currently the SQL endpoint seems to lag. 

The warehouse seems to be more versatile, though careful attention needs to be given to its design. 

The Eventhouse opens the door to a wide range of time-based scenarios, though it will be interesting how developers cope with its lack of functionality in some areas. 

Fabric SQL databases are a new addition, and hopefully they'll allow considering a wide range of OLTP scenarios. 

Power BI datamarts have been in preview for a couple of years.

References:
[1] Microsoft Fabric (2024) Microsoft Fabric decision guide: choose a data store [link]

[2] Reitse's blog (2024) Testing Microsoft Fabric Capacity: Data Warehouse vs Lakehouse Performance [link]