💠🏭🗒️Microsoft Fabric: Polaris SQL Pool [Notes]

Disclaimer: This is work in progress intended to consolidate information from various sources and may deviate from them. Please consult the sources for the exact content!

Unfortunately, besides the references papers, there's almost no material that could be used to enhance the understanding of the concepts presented. 

Last updated: 26-Mar-2025

Read and Write Operations in Polaris [2]

[Microsoft Fabric] Polaris SQL Pool

• {def} distributed SQL query engine that powers Microsoft Fabric's data warehousing capabilities
• designed to unify data warehousing and big data workloads while separating compute and state for seamless cloud-native operations
• based on a robust DCP 
• designed to execute read-only queries in a scalable, dynamic and fault-tolerant way [1]
• a highly-available micro-service architecture with well-defined responsibilities [2]
• data and query processing is packaged into units (aka tasks) 
• can be readily moved across compute nodes and re-started at the task level
• widely-partitioned data with a flexible distribution model [2]
• a task-level "workflow-DAG" that is novel in spanning multiple queries [2]
• a framework for fine-grained monitoring and flexible scheduling of tasks [2]
• {component} SQL Server Front End (SQL-FE)
• responsible for 
• compilation
• authorization
• authentication
• metadata
• used by the compiler to 
• {operation} generate the search space (aka MEMO) for incoming queries
• {operation} bind metadata to data cells
• leveraged to ensure the durability of the transaction manifests at commit [2]
• only transactions that successfully commit need to be actively tracked to ensure consistency [2]
• any manifests and data associated with aborted transactions are systematically garbage-collected from OneLake through specialized system tasks [2]
• {component} SQL Server Backend (SQL-BE)
• used to perform write operations on the LST [2]
• inserting data into a LST creates a set of Parquet files that are then recorded in the transaction manifest [2]
• a transaction is represented by a single manifest file that is modified concurrently by (one or more) SQL BEs [2]
• SQL BE leverages the Block Blob API provided by ADLS to coordinate the concurrent writes  [2]
• each SQL BE instance serializes the information about the actions it performed, either adding a Parquet file or removing it [2]
• the serialized information is then uploaded as a block to the manifest file
• uploading the block does not yet make any visible changes to the file [2]
• each block is identified by a unique ID generated on the writing SQL BE [2]
• after completion, each SQL BE returns the ID of the block(s) it wrote to the Polaris DCP [2]
• the block IDs are then aggregated by the Polaris DCP and returned to the SQL FE as the result of the query [2]
• the SQL FE further aggregates the block IDs and issues a Commit Block operation against storage with the aggregated block IDs [2]
• at this point, the changes to the file on storage will become effective [2]
• changes to the manifest file are not visible until the Commit operation on the SQL FE
• the Polaris DCP can freely restart any part of the operation in case there is a failure in the node topology [2]
• the IDs of any blocks written by previous attempts are not included in the final list of block IDs and are discarded by storage [2]
• [read operations] SQL BE is responsible for reconstructing the table snapshot based on the set of manifest files managed in the SQL FE
• the result is the set of Parquet data files and deletion vectors that represent the snapshot of the table [2]
• queries over these are processed by the SQL Server query execution engine [2]
• the reconstructed state is cached in memory and organized in such a way that the table state can be efficiently reconstructed as of any point in time [2]
• enables the cache to be used by different operations operating on different snapshots of the table [2]
• enables the cache to be incrementally updated as new transactions commit [2]
• {feature} supports explicit user transactions
• can execute multiple statements within the same transaction in a consistent way
• the manifest file associated with the current transaction captures all the (reconciled) changes performed by the transaction [2]
• changes performed by prior statements in the current transaction need to be visible to any subsequent statement inside the transaction (but not outside of the transaction) [2]
• [multi-statement transactions] in addition to the committed set of manifest files, the SQL BE reads the manifest file of the current transaction and then overlays these changes on the committed manifests [1]
• {write operations} the behavior of the SQL BE depends on the type of the operation.
• insert operations 
• only add new data and have no dependency on previous changes [2]
• the SQL BE can serialize the metadata blocks holding information about the newly created data files just like before [2]
• the SQL FE, instead of committing only the IDs of the blocks written by the current operation, will instead append them to the list of previously committed blocks
• ⇐ effectively appends the data to the manifest file [2]
• {update|delete operations} 
• handled differently 
• ⇐ since they can potentially further modify data already modified by a prior statement in the same transaction [2]
• e.g. an update operation can be followed by another update operation touching the same rows
• the final transaction manifest should not contain any information about the parts from the first update that were made obsolete by the second update [2]
• SQL BE leverages the partition assignment from the Polaris DCP to perform a distributed rewrite of the transaction manifest to reconcile the actions of the current operation with the actions recorded by the previous operation [2]
• the resulting block IDs are sent again to the SQL FE where the manifest file is committed using the (rewritten) block IDs [2]
• {concept} Distributed Query Processor (DQP)
• responsible for 
• distributed query optimization
• distributed query execution
• query execution topology management
• {concept} Workload Management (WLM)
•  consists of a set of compute servers that are, simply, an abstraction of a host provided by the compute fabric, each with a dedicated set of resources (disk, CPU and memory) [2]
• each compute server runs two micro-services
• {service} Execution Service (ES) 
• responsible for tracking the life span of tasks assigned to a compute container by the DQP [2]
• {service} SQL Server instance
• used as the back-bone for execution of the template query for a given task  [2]
• ⇐ holds a cache on top of local SSDs 
• in addition to in-memory caching of hot data
• data can be transferred from one compute server to another
• via dedicated data channels
• the data channel is also used by the compute servers to send results to the SQL FE that returns the results to the user [2]
• the life cycle of a query is tracked via control flow channels from the SQL FE to the DQP, and the DQP to the ES [2]
• {concept} cell data abstraction
• the key building block that enables to abstract data stores
• abstracts DQP from the underlying store [1]
• any dataset can be mapped to a collection of cells [1]
• allows distributing query processing over data in diverse formats [1]
• tailored for vectorized processing when the data is stored in columnar formats [1] 
• further improves relational query performance
• 2-dimenstional
• distributions (data alignment)
• partitions (data pruning)
• each cell is self-contained with its own statistics [1]
• used for both global and local QO [1]
• cells can be grouped physically in storage [1]
• queries can selectively reference either cell dimension or even individual cells depending on predicates and type of operations present in the query [1]
• {concept} distributed query processing (DQP) framework
• operates at the cell level 
• agnostic to the details of the data within a cell
• data extraction from a cell is the responsibility of the (single node) query execution engine, which is primarily SQL Server, and is extensible for new data types [1], [2]
• {concept} dataset
• logically abstracted as a collection of cells [1] 
• can be arbitrarily assigned to compute nodes to achieve parallelism [1]
• uniformly distributed across a large number of cells 
• [scale-out processing] each dataset must be distributed across thousands of buckets or subsets of data objects,
•  such that they can be processed in parallel across nodes
• {concept} session
• supports a spectrum of consumption models, ranging from serverless ad-hoc queries to long-standing pools or clusters [1]
• all data are accessible from any session [1]
• multiple sessions can access all underlying data concurrently  [1]
• {concept} Physical Metadata layer
• new layer introduced in the SQL Server storage engine [2]

See also: Polaris

Previous Post <<||>> Next Post

References:
[1] Josep Aguilar-Saborit et al (2020) POLARIS: The Distributed SQL Engine in Azure Synapse, Proceedings of the VLDB Endowment PVLDB 13(12) [link]
[2] Josep Aguilar-Saborit et al (2024), Extending Polaris to Support Transactions [link]
[3] Gjnana P Duvvuri (2024) Microsoft Fabric Warehouse Deep Dive into Polaris Analytic Engine [link]

Resources:

[R1] Microsoft Learn (2025) Fabric: What's new in Microsoft Fabric? [link]
[R2] Patrick Pichler (2023) Data Warehouse (Polaris) vs. Data Lakehouse (Spark) in Microsoft Fabric [link]
[R3] Tiago Balabuch (2023) Microsoft Fabric Data Warehouse - The Polaris engine [link]

Acronyms:

CPU - Central Processing Unit
DAG - Directed Acyclic Graph
DB - Database
DCP - Distributed Computation Platform 
DQP - Distributed Query Processing 

DWH - Data Warehouses 
ES - Execution Service
LST - Log-Structured Table

SQL BE - SQL Backend

SQL FE - SQL Frontend
SSD - Solid State Disk
WAL - Write-Ahead Log

WLM - Workload Management

🧊💫Data Warehousing: Architecture (Part VI: Building a Data Lakehouse for Dynamics 365 Environments with Serverless SQL Pool)

Data Warehousing Series

One of the major limitations of Microsoft Dynamics 365 is the lack of direct access to the production databases for reporting purposes via standard reporting or ETL/ELT tools. Of course, one can attempt to use OData-based solutions though they don't scale with the data volume and imply further challenges. 

At the beginning, Microsoft attempted to address this limitation by allowing the export of data entities into customer's Azure SQL database, feature known as bring your own database (BYOD). Highly dependent on batch jobs, the feature doesn't support real-time synchronization and composite entities, and is dependent on the BYOD's database capacity, the scalability after a certain point becoming a bottleneck.

Then Microsoft started to work on two solutions for synchronizing the Dynamics 365 data in near-real time (cca. 10-30 minutes) to the Data Lake: the Export to Data Lake add-in (*), respectively the Azure Synapse Link for Dataverse with Azure Data Lake. The former allows the synchronization of the tables from Finance & Operations (doesn't work for CRM) to files that reflect the database model from the source. In exchange, the latter allows the synchronization of data entities to similar structures, and probably will support tables as well. Because the service works via Dataverse it supports also the synchronization of CRM data. 

The below diagram depicts the flow of data from the D365 environments to the Data Lake, the arrow indicating the direction of the flow. One arrow could be drawn between Dynamics 365 Finance & Operations and the Azure Link for Datavetse service, though one may choose to use only the Export to Data Lake add-in given that a data model based on the tables offers more flexibility (in the detriment of effort though). Data from other systems can be exported via pipelines to the Data Lake to provide an integrated and complete view of the business. 

Serverless Data Lakehouse

Once the data available in the Delta Lake, it can be consumed directly by standard and Power BI paginated reports, however building the data model will involve considerable effort and logic duplication between reports. Therefore, it makes sense to prepare the data upfront in the Data Lake, when possible, and here the serverless SQL pool can help building an enterprise data model. The former approach can still be used for rapid prototyping or data discovery. 

The serverless SQL Server pool is a stateless SQL-based distributed data processing query service over Azure data lake for large-scale data and computational functions. Even if it doesn't support standard tables, it allows to make the data from the Data Lakes files available for processing via external tables, a mechanism that maps files' structure to an entity that can be queried like a normal view (though it supports only read operations). 

Further on, the enterprise data model can be built like in a normal Data Warehouse via the supported objects (views, stored procedures and table-valued functions). These objects can be called from standard and Power BI paginated reports, the queries being processed at runtime anew, which might result occasionally in poor performance. However, the architecture is supposed to scale automatically as needed.

If further performance is needed, parts of the logic or the end-result can be exported to the Data Lake and here the Medallion Architecture should be considered when appropriate. Upon case, further processing might be needed to handle the limitations of the serverless SQL pool (e.g.flattening hierarchies, handling data quality issues).

One can go around the lack of standard table support needed especially for value mappings by storing the respective data as files and/or occasionally by misusing views, respectively by generating Spark tables via the Spark pool. 

Note:
(*) Existing customers have until 1-Nov-2024 to transition from Export to Data lake to Synapse link. Microsoft advises new customers to use Synapse Link.

Previous Post <<||>> Next Post

🧊Data Warehousing: Architecture V (Dynamics 365, the Data Lakehouse and the Medallion Architecture)

Data Warehousing Series

An IT architecture is built and functions under a set of constraints that derive from architecture’s components. Usually, if we want flexibility or to change something in one area, this might have an impact in another area. This rule applies to the usage of the medallion architecture as well! 

In Data Warehousing the medallion architecture considers a multilayered approach in building a single source of truth, each layer denoting the quality of data stored in the lakehouse [1]. For the moment are defined 3 layers - bronze for raw data, silver for validated data, and gold for enriched data. The concept seems sound considering that a Data Lake contains all types of raw data of different quality that needs to be validated and prepared for reporting or other purposes.

On the other side there are systems like Dynamics 365 that synchronize the data in near-real-time to the Data Lake through various mechanisms at table and/or data entity level (think of data entities as views on top of other tables or views). The databases behind are relational and in theory the data should be of proper quality as needed by business.

The greatest benefit of serverless SQL pool is that it can be used to build near-real-time data analytics solutions on top of the files existing in the Data Lake and the mechanism is quite simple. On top of such files are built external tables in serverless SQL pool, tables that reflect the data model from the source systems. The external tables can be called as any other tables from the various database objects (views, stored procedures and table-valued functions). Thus, can be built an enterprise data model with dimensions, fact-like and mart-like entities on top of the synchronized filed from the Data Lake. The Data Lakehouse (= Data Warehouse + Data Lake) thus created can be used for (enterprise) reporting and other purposes.

As long as there are no special requirements for data processing (e.g. flattening hierarchies, complex data processing, high-performance, data cleaning) this approach allows to report the data from the data sources in near-real time (10-30 minutes), which can prove to be useful for operational and tactical reporting. Tapping into this model via standard Power BI and paginated reports is quite easy. 

Now, if it's to use the data medallion approach and rely on pipelines to process the data, unless one is able to process the data in near-real-time or something compared with it, a considerable delay will be introduced, delay that can span from a couple of hours to one day. It's also true that having the data prepared as needed by the reports can increase the performance considerably as compared to processing the logic at runtime. There are advantages and disadvantages to both approaches. 

Probably, the most important scenario that needs to be handled is that of integrating the data from different sources. If unique mappings between values exist, unique references are available in one system to the records from the other system, respectively when a unique logic can be identified, the data integration can be handled in serverless SQL pool.

Unfortunately, when compared to on-premise or Azure SQL functionality, the serverless SQL pool has important constraints - it's not possible to use scalar UDFs, tables, recursive CTEs, etc. So, one needs to work around these limitations and in some cases use the Spark pool or pipelines. So, at least for exceptions and maybe for strategic reporting a medallion architecture can make sense and be used in parallel. However, imposing it on all the data can reduce flexibility!

Bottom line: consider the architecture against your requirements!

Previous Post <<||>>> Next Post

[1] What is the medallion lakehouse architecture?
https://learn.microsoft.com/en-us/azure/databricks/lakehouse/medallion

🧊Data Warehousing: Architecture (Part IV: Building a Modern Data Warehouse with Azure Synapse)

Introduction

When building a data warehouse (DWH) several key words or derivatives of them appear in requirements: secure, flexible, simple, scalable, reliable, performant, non-redundant, modern, automated, real-timed, etc. As it proves in practice, all these requirements are sometimes challenging to address with the increased complexity of the architecture chosen. There are so many technologies on the DWH market promising all these at low costs, low effort and high ROI, though DWH projects continue to fail addressing the business and technical requirements.

On a basic level for building a DWH is needed a data storage layer and an ETL (Extract, Transfer, Load) tool responsible for the data movement between the various source systems and DWH, and eventually within the DWH itself. After that, each technology added to the landscape tends to increase the overall complexity (and should be regarded with a critical eye in what concerns the advantages and disadvantages).

A Reference Architecture

When building a DWH or a data migration solution, which has many of the characteristics of a DWH, from the many designs, I prefer to keep things as simple as possible.  An approach based on a performant database engine like SQL Server as storage layer and SSIS (SQL Server Integration Services) as ETL proved to be the best choice until now, allowing to address most of the technical requirements by design. Then come the choices on how and where to import and transform the data, at what level of granularity, on how the semantic layer is built, how the data are accessed, etc.

Being able to pull (see extract subprocess) the data from the data sources on a need by basis offers the most flexible approach, however there are cases in which the direct access to source data is not possible, having to rely on a push approach, where data are dumped regularly to a given location (e.g. FTP folder structure), following to be picked up as needed. It's actually a hybrid between a push and pull, because a fully push approach would mean pushing the data directly to the DWH, which can be also acceptable, though might offer lower control on data's movement and involve a few other challenges (e.g. permissions, concurrency). 

Data can be prepared for the DWH in the source systems (e.g. exposed via data objects or API calls), anywhere in between via ETL-based transformations (see transform subprocess) or directly in the DWH. I prefer importing the data (see load subprocess) 1:1 without any transformations from the various sources via SSIS (or similar technologies) into a set of tables that designated the staging area. It's true that in this way the ETL technology is used to a minimum, though unless there's a major benefit to use it for data transformations, using DWH's capabilities and SQL for data processing can provide better performance and flexibility. 

Besides the selection of the columns in scope (typically columns with meaningful values), it's important not to do any transformations in the extraction layer because the data is imported faster (eventually using fast load options as in SSIS) and it assures a basis for troubleshooting (as the data don't change between loads). Some filters can be applied only when the volume of data is high, and the subset of the data could be identified clearly (e.g. when data are partitioned based on a key like business unit, legal entity or creation date).

For better traceability, the staging schemas can reflect the systems they come from, the tables and the columns should have the same names, respectively same data types. On such tables no constraints are applied and no indexes are needed. They can be constructed however on the production tables (aka base tables) - copy of the tables from production. 

Some DWH architects try replicating the constraints from the source systems and/or add more constraints on top to define the various business rules. Rigor is good in some scenarios, though it can involve a considerable effort and it might be challenging to keep over time, especially when considering the impact of big data on DWH architectures. Instead of using constraints, building a set of SQL scripts that pinpoint the issues as reports allow more flexibility with the risk of having inconsistencies running wild through the reports. The data should be cleaned in the source system and not possible then properly addressed in the DWH. Applying constraints will make the data unavailable for reporting until data are corrected, while being more permissive would allow dirty data. Thus, either case has advantages or disadvantages, though the latter seems to be more appropriate. 

Indexes on the production schema should reflect the characteristics of the queries run on the data and shouldn't replicate the indexes from the source environments, even if some overlaps might exist. In practice, dropping the non-clustered indexes on the production tables before loading the data from staging, and recreating them afterwards proves to provide faster loading (see load optimization techniques). 

The production tables are used for building a "semantic" data model or something similar. Several levels of views, table-valued functions and/or indexed/materialized views allows building the dimensions and facts tables, the latter incorporating the business logic needed by the reports. Upon case, stored-procedures, physical or temporary tables, table variables can be used to prepare the data, though they tend to break the "free" flow of data as steps in-between need to be run. On the other side, in certain scenarios their use is unavoidable. 

The first level of views (aka base views) is based on the base tables without any joins, though they include only the fields in use (needed by the business) ordered and "grouped" together based on their importance or certain characteristics. The views can include conversions of data types, translations of codes into meaningful values, and quite seldom filters on the data. Based on these "base" views the second level is built, which attempts to define the dimension and fact tables at the lowest granularity. These views include joins between tables coming from the same or different systems, respectively mappings of values defined in tables, and whatever it takes to build such entities. However, transformations on individual fields are pushed, when possible, to the lower level to minimize logic redundancy. From similar reasons, the logic could be broken down over two or more "helper" views when visible benefits could be obtained from it (e.g troubleshooting, reuse, maintenance). It's important to balance between creating too many helper views and encapsulating too much logic in a view. 

One of the design principles used in building the entities is to minimize the redundance of the fields used, ideally without having columns duplicated between entities at this level. This would facilitate the traceability of columns to the source tables within the "semantic" layer (typically in the detriment of a few more joins). In practice, one is forced to replicate some columns to simplify some parts of the logic. 

Further views can be built based on the dimension and fact entities to define the logic needed by the reports. Only these objects are used and no direct reference to the "base" tables or views are made. Moreover, to offer better performance when the views can be materialized or, when there's an important benefit, physically saved as table (e.g. having multiple indexes for different scenarios). It's the case of entities with considerable data volume called over and over. 

This approach of building the entities is usually flexible enough to address most of the reporting requirements, independently whether the technical solution has the characteristics of a DWH, data mart or data migration layer. Moreover, the overall architectural approach can be used on-premise as well in cloud architectures, where Azure SQL Server and ADF (Azure Data Factory) provide similar capabilities. Compared with standard SQL Server, some features might not be available, while other features might bring further benefits, though the gaps should be neglectable.

Data Management topics like Master Data Management (MDM), Data Quality Management (DQM) and/or Metadata Management can be addressed as well by using third-party tools or tools from the Microsoft stack - Master Data Services (MDS) and Data Quality Services (DQS) in combination with SSIS help addressing a wide range of scenarios - however these are optional. 

Moving to the Cloud

Within the context of big data, characterized by (high/variable) volume, value, variety, velocity, veracity, and further less important V's, the before technical requirements still apply, however within a cloud environment the overall architecture becomes more complex. Each component becomes a service. There are thus various services for data ingestion, storage, processing, sharing, collaboration, etc. The way data are processed involves also several important transformations: ETL becomes ELT, FTP and local storage by Data Lakes, data packages by data pipelines, stateful by stateless, SMP (Symmetric Multi-Processing) by MPP (Massive Parallel Processing), and so on.

As file storage is less expensive than database storage, there's an increasing trend of dumping business critical data into the Data Lake via data pipelines or features like Link to Data Lake or Export to Data Lake (*), which synchronize the data between source systems and Data Lake in near real-time at table or entity level. Either saved as csv, parquet, delta lake or any other standard file format, in single files or partitions, the data can be used directly or indirectly for analytics.

Cloud-native warehouses allow addressing topics like scalability, elasticity, fault-tolerance and performance by design, though further challenges appear as compute needs to be decoupled from storage, the workloads need to be estimated for assuring the performance, data may be distributed across data centers spanning geographies, the infrastructure is exposed to attacks, etc. 

Azure Synapse

If one wants to take advantage of the MPP architecture's power, Microsoft provides an analytical architecture based on Azure Synapse, an analytics service that brings together data integration, enterprise DWH, and big data analytics. Besides two types of SQL-based data processing services  (dedicated vs serverless SQL pools) it comes also with a Spark pool for in-memory cluster computing.

A DWH based on Azure Synapse is not that different from the reference architecture described above for an on-premise solution. Actually, a DWH based on a dedicated SQL pool (aka a physical data warehouse) involves the same steps mentioned above. 

The data can be imported via ETL/ELT pipelines in the DWH, though there are also mechanisms for consuming the data directly from the files stored in the Data Lake or Azure storage. CETAS (aka Create External Table as Select) can be defined on top of the data files, the external tables acting as "staging" or "base" tables in the architecture described above. When using a dedicated SQL pool it makes sense to use the CETAS as "staging" tables, the processed data following to be dumped to "optimized" physical tables for consumption and refreshed periodically. However, when this happens the near real-time character of data is lost. Using the CETAs as base tables would keep this characteristic as long the data isn't saved physically in tables or files, maybe in the detriment of performance.

Using a dedicated SQL pool for direct reporting can become expensive as the pool needs to be available at least during business hours for incoming user requests, or at least for importing the data and refreshing the datasets. When using the CETAS as a base table, a serverless (aka on-demand) SQL pool, which uses a per-pay-use billing model could prove to be more cost-effective and flexible in many scenarios. By design, it helps to keep the near real-time character of the data. Moreover, even if the data are actually moved from the source tables into the Data Lake, this architecture has the characteristics of a logical data warehouse:

Unfortunately, unless one uses Spark tables, misuses views or adds an Azure SQL database to the architecture, there are no physical tables or materialized views in a serverless SQL pool. There's still the option to use data pipelines for regullarly exporting intermediary data to files (incl. over partitions or folders), even if this involves more overhead as it's not possible to export data over SQL syntax to files more than once (though this might change in the future). For certain scenario it could be useful to store data in a Azure SQL Server or similar database, including a dedicated SQL pool. 

Choosing between serverless and dedicated SQL pool is not an exclusive choice, both or all 3 types of pools (if we consider also the Spark pool) can be used in the architecture for addressing specific challenges, especially when we consider that there are important differences between the features available in each of the pools. Moreover, one can start the PoC based on the serverless SQL pool and when the solution became mature enough and used in all enterprise, parts of the logic or all of it can be migrated to a dedicated SQL pool. This would allow to save costs at the beginning in the detriment of further effort later. 

Talking about the physical storage, data engineers recommend defining within a Data Lake several layers (aka regions, zones) labeled as bronze, silver and gold (and probably platinum will join the club anytime soon). The bronze layer refers to the raw data available in the Data Lake, including the files on which the initial CETAS are defined upon. The silver refers to transformed, cleaned, enriched and integrated data, data resulting from the second layer of views described above. The gold layer refers to the data to which business logic was applied and prepared for consumption, data resulting from the final layer of views. Of course, data pipelines can be used to prepare the data at these stages, though a view-based approach offers more flexibility, are easier to troubleshoot, manage and reuse than data pipelines.

Ideally the gold data should involve no or minimal further transformation before reaching the users, though that's not realistic. Building a DWH takes a considerable time and the business can't usually wait until everything is in place. Therefore, reports based on DWH will continue to coexist with reports directly accessing the source data, which will lead to controversies. Enforcing a single source of truth will help to minimize the gap, though will not eliminate it completely. 

Closing Notes

These are just outlines of a minimal reference architecture. There's more to consider, as there are several alternatives (see [1] [2] [3] [4]) for each of the steps considered in here, each technology, new features or mechanisms opening new opportunities. The advantages and disadvantages should be always considered against the business needs and requirements. One approach, even if recommended, might not work for all, though unless there's an important requirement or an opportunity associated with an additional technology, deviating from reference architectures might not be such a good idea afterall.

Note:

(*) Existing customers have until 1-Nov-2024 to transition from Export to Data lake to Synapse link. Microsoft advises new customers to use Synapse Link. 

Previous Post <<||>> Next Post

Resources:

[1] Microsoft Learn (2022) Modern data warehouse for small and medium business (link)

[2] Microsoft Learn (2022) Data warehousing and analytics (link)

[3] Microsoft Learn (2022) Enterprise business intelligence (link)

[4] Microsoft Learn (2022) Serverless Modern Data Warehouse Sample using Azure Synapse Analytics and Power BI (link)

[5] Coursera (2023) Data Warehousing with Microsoft Azure Synapse Analytics (link) [course, free to audit]

[6] SQLBits (2020) Mahesh Balija's Building Modern Data Warehouse with Azure Synapse Analytics (link)

[7] Matt How (2020) The Modern Data Warehouse in Azure: Building with Speed and Agility on Microsoft’s Cloud Platform (Amazon)

[8] James Serra's blog (2022) Data lake architecture (link)

[9] SQL Stijn (2022) SQL Building a Modern Lakehouse Data Warehouse with Azure Synapse Analytics: Moving your Database to the lake (link)

[10] Solliance (2022) Azure Synapse Analytics Workshop 400 (link) [GitHub repository]

💎🏭SQL Reloaded: Automatic Statistics Creation & Dropping for CETAS based on CSV File Format in Serverless SQL Pool

Introduction

The serverless SQL pool query optimizer uses statistics to calculate and compare the cost of various query plans, and then choose the plan with the lowest cost. Automatic creation of statistics is turned on for parquet file format, though for CSV file format statistics will be automatically created only when OPENROWSET is used. This means that when creating CETAS based on CSV the statistics need to be created manually. 

This would be one more reason for holding the files in the Data Lake as parquet files. On the other side there are also many files already available in CSV format, respectively technoloqies that allows exporting data only/still as CSV. Moreover, transforming the files as parquet is not always technically feasible.

Using OPENROWSET could also help, though does it make sense to use a different mechanismus for the CSV file format? In some scenarios will do. I prefer to have a unitary design, when possible. Moreover, even if some columns are not needed, they can still be useful for certain scenarios (e.g. troubleshooting, reevaluating their use, etc.). 

There are files, especially the ones coming from ERPs (Enterprise Resource Planning) or similar systems, which have even a few hundred columns (on average between 50 and 100 columns). Manually creating  the statistics for the respective tables will cost lot of time and effort. To automate the process there are mainly three choices:
(1) Creating statistics for all the columns for a given set of tables (e.g. for a given schema).
(2) Finding a way to automatically identify the columns which are actually used.
(3) Storing the list of tables and columns on which statistics should be build (however the list needs to be maintained manually). 

Fortunately, (1) can be solved relatively easy, based on the available table metadata, however it's not the best solution, as lot of statistics will be unnecessarily created. (2) is possible under certain architectures or additional effort. (3) takes time, though it's also an approachable solution.

What do we need?

For building the solution, we need table and statistics metadata, and the good news is that the old SQL Server queries still work. To minimize code's repetition, it makes sense to encapsulate the logic in views. For table metadata one can use the sys.objects DMV as is more general (one can replace sys.objects with sys.tables to focus only on tables):

-- drop the view (for cleaning)
-- DROP VIEW IF EXISTS dbo.vAdminObjectColumns

-- create view
CREATE OR ALTER VIEW dbo.vObjectColumns
AS 
-- object-based column metadata
SELECT sch.name + '.' + obj.name two_part_name
, sch.Name schema_name
, obj.name object_name
, col.name column_name
, obj.type
, CASE 
	WHEN col.is_ansi_padded = 1 and LEFT(udt.name , 1) = 'n' THEN col.max_length/2
	ELSE col.max_length
  END max_length
, col.precision 
, col.scale
, col.is_nullable 
, col.is_identity
, col.object_id
, col.column_id
, udt.name as data_type
, col.collation_name
, ROW_NUMBER() OVER(PARTITION BY col.object_id  ORDER BY col.column_id) ranking
FROM sys.columns col
     JOIN sys.types udt
       on col.user_type_id= udt.user_type_id 
	 JOIN sys.objects obj
	   ON col.object_id = obj.object_id
	 JOIN sys.schemas as sch
	   on sch.schema_id = obj.schema_id

-- testing the view
SELECT obc.*
FROM dbo.vObjectColumns obc
WHERE obc.object_name LIKE '<table name>%'
  AND obc.schema_name = 'CRM'
  AND obc.type = 'U'
ORDER BY obc.two_part_name
, Ranking

The view can be used also as basis for getting the defined stats:

-- drop the view (for cleaning)
-- DROP VIEW IF EXISTS dbo.vAdminObjectStats

-- create view 
CREATE OR ALTER VIEW dbo.vObjectStats
AS
-- object-based column statistics
SELECT obc.two_part_name + '.' + QuoteName(stt.name) three_part_name
, obc.two_part_name
, obc.schema_name
, obc.object_name
, obc.column_name
, stt.name stats_name
, STATS_DATE(stt.[object_id], stt.stats_id) AS last_updated
, stt.auto_created
, stt.user_created
, stt.no_recompute
, stt.has_filter 
, stt.filter_definition
, stt.is_temporary 
, stt.is_incremental 
, stt.auto_drop 
, stt.stats_generation_method_desc
, stt.[object_id]
, obc.type 
, stt.stats_id
, stc.stats_column_id
, stc.column_id
FROM dbo.vObjectColumns obc
     LEFT JOIN sys.stats_columns stc 
	   ON stc.object_id = obc.object_id
	  AND stc.column_id = obc.column_id 
          LEFT JOIN sys.stats stt
            ON stc.[object_id] = stt.[object_id] 
           AND stc.stats_id = stt.stats_id

-- testing the view 
SELECT *
FROM dbo.vObjectStats obs
WHERE (obs.auto_created = 1 OR obs.user_created = 1)
  AND obs.type = 'U'
  AND obs.object_name = '<table name>'
ORDER BY obs.two_part_name
, obs.column_id

Now we have a basis for the next step. However, before using the stored procedure define below, one should use the last query and check whether statistics were defined before on a table. Use for testing also a table for which you know that statistics are available.

Create Statistics

The code below is based on a similar stored procedure available in the Microsoft documentation (see [1]). It uses a table's column metadata, stores them in a temporary table and then looks through each record, create the DDL script and runs it:

-- drop procedure (for cleaning)
--DROP PROCEDURE dbo.pCreateStatistics

-- create stored procedure
CREATE OR ALTER PROCEDURE dbo.pCreateStatistics
(   @schema_name nvarchar(50)
,   @table_name nvarchar(100)
)
AS
-- creates statistics for serverless SQL pool
BEGIN
	DECLARE @query as nvarchar(1000) = ''
	DECLARE @index int = 1, @nr_records int = 0

	-- drop temporary table if it exists 
	DROP TABLE IF EXISTS #stats_ddl;

	-- create temporary table 
	CREATE TABLE #stats_ddl( 
	  schema_name nvarchar(50)
	, table_name nvarchar(128)
	, column_name nvarchar(128)
	, ranking int
	);

	-- fill table
	INSERT INTO #stats_ddl
	SELECT obc.schema_name
	, obc.object_name
	, obc.column_name 
	, ROW_NUMBER() OVER(ORDER BY obc.schema_name, obc.object_name) ranking
	FROM dbo.vObjectColumns obc
	WHERE obc.type = 'U' -- tables
	  AND IsNull(@schema_name, obc.schema_name) = obc.schema_name 
	  AND IsNull(@table_name, obc.object_name) = obc.object_name

	SET @nr_records = (SELECT COUNT(*) FROM #stats_ddl)

	WHILE @index <= @nr_records
	BEGIN
		SET @query = (SELECT 'CREATE STATISTICS '+ QUOTENAME('stat_' + schema_name + '_' + table_name + '_' + column_name) + ' ON '+ QUOTENAME(schema_name) + '.' + QUOTENAME(table_name) + '(' + QUOTENAME(column_name) + ')' 
			   FROM #stats_ddl ddl
			   WHERE ranking = @index);

		BEGIN TRY
		        -- execute ddl
			EXEC sp_executesql @query;
		END TRY
		BEGIN CATCH
			SELECT 'create failed for ' + @query;
		END CATCH

		SET @index+=1;
	END

	DROP TABLE #stats_ddl;
END


-- test stored procedure (various scenario)
EXEC dbo.pCreateStatistics '<schema name>', '<table name>' -- based on schema & table
EXEC dbo.pCreateStatistics '<schema name>', NULL -- based on a schema
EXEC dbo.pCreateStatistics NULL, '<table name>' -- based on a table

Notes:
IMPORTANT!!! I recommend testing the stored procedure in a test environment first for a few tables and not for a whole schema. If there are too many tables, this will take time.

Please note that rerunning the stored procedure without deleting previously the statitics on the tables in scope will make the procedure raise failures for each column (behavior by design), though the error messages can be surpressed by commenting the code, if needed. One can introduce further validation, e.g. considering only the columns which don't have a statistic define on them.

Further Steps?

What can we do to improve the code? It would be great if we could find a way to identify the columns which are used in the queries. It is possible to retrieve the queries run in serverless SQL pool, however identifying the tables and columns from there or a similar source is not a straightforward solution. 

The design of views based on the external tables can help in the process! I prefer to build on top of the external tables a first level of views (aka "base views") that include only the fields in use (needed by the business) ordered and "grouped" together based on their importance or certain characteristics. The views are based solely on the external table and thus contain no joins. They can include conversions of data types, translations of codes into meaningful values, and quite seldom filters on the data. However, for traceability the name of the columns don't change! This means that if view's name is easily identifiable based on external table's name, we could check view's columns against the ones of the external table and create statistics only for the respective columns. Using a unique prefix (e.g. "v") to derive views' name from tables' name would do the trick.

To do that, we need to create a view that reflects the dependencies between objects (we'll be interested only in external tables vs views dependencies):

-- drop view (for cleaning)
-- DROP VIEW IF EXISTS dbo.vObjectsReferenced

-- create view
CREATE OR ALTER VIEW dbo.vObjectsReferenced
AS 
-- retrieving the objects referenced 
SELECT QuoteName(sch.name) + '.' + QuoteName(obj.name) AS two_part_name 
, obj.object_id 
, obj.schema_id 
, sch.name schema_name 
, obj.name object_name 
, obj.type
, QuoteName(scr.name) + '.'+ QuoteName(sed.referenced_entity_name) AS ref_two_part_name 
, obr.object_id ref_object_id
, obj.schema_id ref_schema_id 
, scr.name ref_schema_name 
, obr.name ref_object_name 
, obr.type ref_type
FROM sys.sql_expression_dependencies sed 
     JOIN sys.objects obj
       ON obj.object_id = sed.referencing_id 
	      JOIN sys.schemas as sch
	        ON obj.schema_id = sch.schema_id 
	 JOIN sys.objects obr
	   ON sed.referenced_id = obr.object_id
	      JOIN sys.schemas as scr
	        ON obr.schema_id = scr.schema_id

-- testing the view
SELECT top 10 *
FROM dbo.vObjectsReferenced
WHERE ref_type = 'U'

With this, the query used above to fill the table becomes:

-- fill table query with column selection 
SELECT obc.schema_name
, obc.object_name
, obc.column_name 
, ROW_NUMBER() OVER(ORDER BY obc.schema_name, obc.object_name) ranking
FROM dbo.vObjectColumns obc
WHERE obc.type = 'U' -- tables
	AND IsNull(@schema_name, obc.schema_name) = obc.schema_name 
	AND IsNull(@table_name, obc.object_name) = obc.object_name
	AND EXISTS ( -- select only columns referenced in views
	    SELECT * 
		FROM dbo.vObjectsReferenced obr 
		    JOIN dbo.vAdminObjectColumns obt
			ON obr.object_id = obt.object_id 
		WHERE obt.type = 'V' -- view
		AND obr.object_name  =  'v' + obr.ref_object_name
		AND obc.object_id = obr.ref_object_id
		AND obc.column_name = obt.column_name);

This change will reduce the number of statistics created on average by 50-80%. Of course, there will be also cases in which further statistics need to be added manually. One can use this as input for an analysis of the columns used and store the metadata in a file, do changes to it and base on it statistics' creation. 

Drop Statistics

Dropping the indexes resumes to using the dbo.vObjectStats view created above for the schema and/or table provided as parameter. The logic is similar to statistics' creation:

-- drop stored procedure (for cleaning)
-- DROP PROCEDURE IF EXISTS dbo.pDropStatistics

-- create procedure
CREATE OR ALTER PROCEDURE dbo.pDropStatistics
(   @schema_name nvarchar(50)
,   @table_name nvarchar(128)
)
AS
-- drop statistics for a schema and/or external table in serverless SQL pool
BEGIN

	DECLARE @query as nvarchar(1000) = ''
	DECLARE @index int = 1, @nr_records int = 0

	-- drop temporary table if it exists 
	DROP TABLE IF EXISTS #stats_ddl;

	-- create temporary table 
	CREATE TABLE #stats_ddl( 
	 three_part_name nvarchar(128)
	, ranking int
	);

	-- fill table
	INSERT INTO #stats_ddl
	SELECT obs.three_part_name
	, ROW_NUMBER() OVER(ORDER BY obs.three_part_name) ranking
	FROM dbo.vObjectStats obs
	WHERE obs.type = 'U' -- tables
	  AND IsNull(@schema_name, obs.schema_name) = obs.schema_name 
	  AND IsNull(@table_name, obs.object_name) = obs.object_name

	SET @nr_records = (SELECT COUNT(*) FROM #stats_ddl)

	WHILE @index <= @nr_records
	BEGIN
   
		SET @query = (SELECT 'DROP STATISTICS ' + ddl.three_part_name
			   FROM #stats_ddl ddl
			   WHERE ranking = @index);

		BEGIN TRY
		        -- execute ddl
			EXEC sp_executesql @query;
		END TRY
		BEGIN CATCH
			SELECT 'drop failed for ' + @query;
		END CATCH

		SET @index+=1;
	END

	DROP TABLE #stats_ddl;
END

Note:
IMPORTANT!!! I recommend testing the stored procedure in a test environment first for a few tables and not for a whole schema. If there are too many tables, this will take time.

Closing Thoughts

The solution for statistics' creation is not perfect, though it's a start! It would have been great if such a feature would be provided by Microsoft, and probably they will, given the importance of statistics of identifying an optimal plan. It would be intersting to understand how much statistics help in a distributed environment and what's the volume of data processed for this purpose. 

Please let me know if you found other workarounds for statistics' automation.

Notes:
1) The above objects and queries seem to work also in SQL databases in Microsoft Fabric.

2) For a refresher on statistics see my notes. 

Happy coding!

Previous Post <<||>> Next Post

References:
[1] Microsoft Learn (2022) Statistics in Synapse SQL (link)