The Polaris Papers: Inside the Fabric Data Warehouse
Polaris is the distributed SQL engine behind Fabric Data Warehouse, bringing SQL Server query processing to distributed compute over shared OneLake storage.
Based on the Polaris papers, this session explains how T-SQL in SQL Pools is compiled by the Front End nodes, optimized into distributed physical plans, orchestrated by the Distributed Computing Platform as workflow DAGs, and executed by Back End compute nodes over data cells. We connect that engine design to Parquet storage, Delta Lake log publishing, query decorrelation, and snapshot isolation, and show what it means for data movement, concurrency, plan quality, and performance at scale.
This is a 2x90-minute deep-dive session (level 500).
Slides
The slides will be available on this page after the session.
Benchmarks and query plans
Explore the recorded TPC-H, TPC-DS and DSB timings from the SF100 Fabric Warehouses on F64. All 177 query plans include the SQL, operator details and source XML. The six presentation examples are marked Demo.
Benchmark results · Interactive query plans
Research papers
These are the papers used in the presentation. Start with the Polaris architecture paper, then follow the topics you want to explore further.
Polaris, query planning and transactions
- Access Path Selection in a Relational Database Management System , Selinger et al., SIGMOD 1979. The System R paper that introduced the plan-selection approach and interesting orders discussed in the session.
- Polaris: The Distributed SQL Engine in Azure Synapse , PVLDB 2020.
- Unified Query Optimization in the Fabric Data Warehouse , SIGMOD 2024.
- Extending Polaris to Support Transactions , SIGMOD 2024, author preprint.
- Query Decorrelation in the Fabric Data Warehouse , SIGMOD 2025.
- Bitmap Filtering in the Fabric Data Warehouse , SIGMOD 2026.
- Extensible Query Optimizers in Practice , Foundations and Trends in Databases, 2024.
- Towards Query Optimizer as a Service (QOaaS) in a Unified LakeHouse Ecosystem: Can One QO Rule Them All? , CIDR 2025, author preprint.
- I Can’t Believe It’s Not Yannakakis: Pragmatic Bitmap Filters in Microsoft SQL Server , CIDR 2026.
GPU acceleration
- CoddSpeed: Hardware Accelerated Query Processing in Microsoft Fabric , SIGMOD 2026.
- Rubberband: Memory-Elastic, Skew-Tolerant GPU Hash Joins , ADMS 2026.
- ZipFlow: a Compiler-based Framework to Unleash Compressed Data Movement for Modern GPUs , author preprint, 2026.
- GPU Acceleration of SQL Analytics on Compressed Data , author preprint, PVLDB 2026.
- TQP++: Bridging ML Compilers and Analytical Query Processing on GPUs , PVLDB 2026.
- Scalable GPU Acceleration of Scalar Functions in Analytical Databases: Compilation, Benchmarking, and Optimization , PVLDB 2026.
- Terabyte-Scale Analytics in the Blink of an Eye , PVLDB 19(2), 2025.
Formats and documentation
- Apache Parquet file format .
- Delta Lake transaction log protocol .
- Substrait : relational operators , field references and function declarations .
- Substrait output mapping with emit and binary serialization .
- Custom SQL pools and workload management in Fabric Data Warehouse.
The worked Substrait plan
contains the electricity usage example used in the slides. It describes SELECT account_id FROM EnergyUsage WHERE kwh > 100 with field references, function declarations and the final output mapping. This small example uses whole kWh in signed 64-bit integer fields and the Substrait v0.103.0 schema.
Fabric Data Warehouse behavior
These Microsoft Learn pages cover the product behavior used in the examples and connect it to the mechanisms in the papers.
- Warehouse and SQL analytics endpoints and warehouse architecture .
- Data clustering and V-Order : physical layout and the cost of writing data for repeated reads.
- Memory and SSD caching and result set caching : different forms of reuse during query execution.
- Statistics and performance guidelines : estimates, plan costs, ingestion and parallel execution.
- Query acceleration and its FAQ : eligible SQL work, preview scope, Power BI DirectQuery and billing.
- SQL analytics endpoint performance : metadata sync, scan parallelism and Lakehouse table maintenance.
- Warehouse troubleshooting and query blocking : runtime evidence, table-level conflicts and schema-lock waits.
See you at the next one?
I announce upcoming talks on LinkedIn — that's also where most of the conference chatter happens. Slides and recordings land right here on the speaking page. If you'd rather follow along quietly, the RSS feed has every new post and talk.