🏭🧊🗒️Microsoft Fabric: Partitions in Lakehouses [Notes]

 Disclaimer: This is work in progress intended to consolidate information from various sources for learning purposes. For the latest information please consult the documentation (see the links below)! 

Last updated: 14-Feb-2024

[Microsoft Fabric] Partitions

• {def} a data organization technique used to split a large dataset into smaller, more manageable nonoverlapping subsets (aka partitions, shards) 
• a pattition is defined based on one or more fields
• each partition contains a subset of the data
• each partitions can be stored and processed independently
• {goal} improve performance, scalability, and manageability of large data tables
• {benefit} allows to split large tables into smaller, manageable partitions based on specific criteria [2]
• e.g., date ranges, regions, categories, entities
• {benefit} allows to improve queries' performance as they can target specific partitions [2]
• reduces the amount of data scanned [2]
• improves queries' performance [2]
• {benefit} allows for more efficient data loading [2]
• {benefit} facilitates the management of big tables [2]
• maintenance tasks can be performed on individual partitions  [2]
• obsolete data partitions can be removed with no overhead, adding new partitions on a need basis [2]
• applies to 
• backups
• indexing
• allows optimizing query performance for specific subsets of data
• statistics
• performance can be affected by
• the choice of partition columns for a delta table [1]
• the number and size of partitions of the partition column [1]
• a column with high cardinality (mostly or entirely made of unique values) results in a large number of partitions [1]
• ⇐ negatively impacts performance of the metadata discovery scan for changes [1]
• {recommendation} if the cardinality of a column is high, choose another column for partitioning [1]
• the size of each partition can also affect performance
• {recommendation} use a column that would result in a partition of at least (or close to) 1 GB [1]
• {recommendation} follow the best practices for delta tables maintenance [1]
• a large volume of small-sized parquet files increases the time it takes to sync the changes [1]
• ⇒ leads to large number of parquet files in a delta table due to [1]
• over-partitioning
• partitions with high number of unique values [1]
• {recommendation} choose a partition column that 
• doesn't have a high cardinality [1]
• results in individual partition size of at least 1 GB [1]
• many small files
• batch and streaming data ingestion rates might result in small files
• depends on the frequency and size of changes being written to a lakehouse [1]
• {recommendation} implement regular lakehouse table maintenance [1] 

Previous Post <<||>> Next Post

References:
[1] Microsoft Learn (2024) Microsoft Fabric: SQL analytics endpoint performance considerations [link]

[2] Kenneth A Omorodion (2024) Partitioning Data in Microsoft Fabric to Improve Performance

written [link]

[3] Microsoft Learn (2024) Microsoft Fabric: Loading Fabric Lakehouse Tables with partitions [link]

[4] 

Resources

[R1] Microsoft Learn (2024) Microsoft Fabric: Load data to Lakehouse using partition in a Data pipeline [link]

🏭💠🗒️Microsoft Fabric: SQL Analytics Endpoint [Notes]

Disclaimer: This is work in progress intended to consolidate information from various sources for learning purposes. For the latest information please consult the documentation (see the links below)! 

Last updated: 12-Feb-2024

[Microsoft Fabric] SQL Analytics Endpoint

• {def} a service that listens actively for requests providing a SQL-based experience for lakehouse Delta tables [1]
  • enables to query data in the lakehouse using T-SQL language and TDS protocol [1]
  • created by default for each lakehouses and mirrored databases provisioned in a workspace
  • every lakehouse has one SQL analytics endpoint [1]
  • manages the automatically generated tables so the workspace users can't modify them [1]
• {feature} a background process is responsible for 
• scanning lakehouse for changes [1]
• automatically generates a table in the appropriate schema for every Delta table in the Lakehouse
• tables are created with a minor delay [1]
• data is refreshed automatically
• the amount of time it takes to refresh the table is related to how optimized the Delta tables are [1]
• {alternative} one can manually force a refresh of the automatic metadata scanning in the Fabric portal [1]
• keeping SQL analytics endpoint up-to-date for all the changes committed to lakehouses in a workspace [1]
• {characteristic} transparently managed
• ⇐ the sync process is transparently managed by Microsoft Fabric platform [1]
• when a change is detected in the lakehouse
• a background process updates metadata [1]
• the SQL analytics endpoint reflects the changes committed to lakehouse tables [1]
• {characteristic} low latency
• under normal operating conditions, the lag between a lakehouse and SQL analytics endpoint is less than one minute [1]
• the actual length of time can vary from a few seconds to minutes [1]
• the metadata synchronization is automatically triggered when the SQL Endpoint is opened [3]
• ensures the SQL Analytics Endpoint remains up to date without the need for manual syncs [3]
•  ⇐ provides a seamless experience [3]
• {feature} can be enriched by adding database objects
• schemas
• views
• procedures
• other database objects
• {feature} automatic metadata discovery
• tracks changes committed to lakehouses [1]
• is a single instance per Fabric workspace [1]
• {issue} increased latency for changes to sync between lakehouses and SQL analytics endpoint, it could be due to large number of lakehouses in one workspace [1]
• {resolution} migrate each lakehouse to a separate workspace [1]
• allows automatic metadata discovery to scale [1]
• {issue} changes committed to a lakehouse are not visible in the associated SQL analytics endpoint
• create a new table in the  [1]
• many committed rows may not be visible in the endpoint [1]
• ⇐ may be impacted by ETL processing that generate large volumes of changes
• {recommendation} initiate an on-demand metadata sync, triggered from the SQL query editor Refresh ribbon option [1]
• forces an on-demand metadata sync, rather than waiting on the background metadata sync to finish [1]
• {issue} if there's no maintenance scheduled for the parquet files, this can result in read overhead and this impacts time it takes to sync changes to SQL analytics endpoint [1]
• {recommendation}schedule regular lakehouse table maintenance operations [1]
• {limitation} not all Delta features are understood by the automatic sync process [1]
• {feature} endpoint reprovisioning
• if the initial provisioning attempt fails, users have the option to try again [2]
• ⇐ without the need to create an entirely new lakehouse [2]
• {benefit} empowers users to self-mitigate provisioning issues in convenient way in the UI avoiding the need for complete lakehouse re-creation [2]

Previous Post <<||>> Next Post

References:
[1] Microsoft Learn (2024) Microsoft Fabric: SQL analytics endpoint performance considerations [link]
[2] Microsoft Learn (2024) Microsoft Fabric: What is the SQL analytics endpoint for a lakehouse? [link]
[3] Microsoft Learn (2024) Microsoft Fabric: What’s new in the Fabric SQL Analytics Endpoint? [link]

💎🏭SQL Reloaded: Microsoft Fabric's SQL Databases (Part VII: Things That Don't Work) [new feature]

Microsoft does relatively a good job in documenting what doesn't work in Microsoft Fabric's SQL Databases. There's a good overview available already in the documentation, though beyond this the current post lists my finding while testing the previously written code on this blog,

USE Database

The standard syntax allows to change via USE the database context to the specified database or database snapshot. Unfortunately, this syntax doesn't seem to be supported currently and unfortunately many scripts seem to abuse of it. Thus, the following line of code throws an error:

-- changing the context
USE master;
GO
USE tempdb;

"Msg 40508, Level 16, State 1, Line 1, USE statement is not supported to switch between databases. Use a new connection to connect to a different database"

However, one can use the 3-part naming convention to reference the various objects:

-- sys metadata - retrieving the database files

SELECT *

FROM tempdb.sys.database_files dbf

ORDER BY name;

Even if the tempdb is not listed in the sys.databases table, it's still available for querying, which can prove helpful for troubleshooting. 

DBCC commands 

The documentation warns that some DBCC commands won't work, though in some cases there are also alternatives. For example:

-- clearing the procedure cache via DBCC
DBCC FREEPROCCACHE;

Output:

"Msg 2571, Level 14, State 9, Line 1, User '<user>' does not have permission to run DBCC freeproccache."

Alternatively, one can use the following command, which seems to work:

-- clearing the procedure cash via ALTER
ALTER DATABASE SCOPED CONFIGURATION CLEAR PROCEDURE_CACHE;

CHECKDB, which checks the logical and physical integrity of all the objects in the specified database, can't be used as well:

 

-- Checking the logical and physical integrity of a database
DBCC CHECKDB();

Output:

"Msg 916, Level 14, State 2, Line 1, The server principal "..." is not able to access the database "..." under the current security context."

The same error message is received for CHECKTABLE, utility which checks the integrity of all the pages and structures that make up the table (or indexed view):

-- checking a table's integrity
DBCC CHECKTABLE ('SalesLT.Address');

Output:

"Msg 916, Level 14, State 2, Line 2, The server principal "..." is not able to access the database "..." under the current security context."

A similar error messages is received for SQLPERF, which provides transaction log space usage statistics for all databases:

-- retrieving the LOGSPACE information for all databases
DBCC SQLPERF (LOGSPACE);

Output: 

"Msg 297, Level 16, State 10, Line 1, The user does not have permission to perform this action."

There are however DBCC commands like SHOW_STATISTICS or SHRINKDATABASE which do work. 

 

-- current query optimization statistics
DBCC SHOW_STATISTICS('SalesLT.Address','PK_Address_AddressID');

Output:

Name	Updated	Rows	Rows Sampled	Steps	Density	Average key length	String Index	Filter Expression	Unfiltered Rows	Persisted Sample Percent
PK_Address_AddressID	Dec 21 2024 3:02AM	450	450	197	1	4	NO		450	0

SHRINKDATABASE shrinks the size of the data and log files in the specified database:

-- shrinking database
DBCC SHRINKDATABASE([AdventureWorks01-...]) WITH NO_INFOMSGS;

Update 29-Jan-2025: According to an answer from Ask the Expert session on Fabric Database [3], Microsoft seems to be working in bringing more DBCC features to SQL databases.

Happy coding!

Previous Post <<||>> Next Post

References:
[1] Microsoft Learn (2024) SQL Server: USE <database> [link]
[2] Microsoft Learn (2024) Database console commands [link]
[3] Microsoft Reactor (2025) Ask The Expert - Fabric Edition - Fabric Databases [link]

💎🏭SQL Reloaded: Microsoft Fabric's SQL Databases (Part VI: Index Usage Analysis) [new feature]

There are several system dynamic management views (DMV) available in SQL Server, Azure SQL Server and now in SQL databases that allow to gather more information about indexes' fragmentation and usage. Let's look at the most important information available based on the indexes create in the previous posts. As the data were probably purged from the views, it's needed to run first the select queries based on the SalesLT.Product from the previous post. This step is important, otherwise the DMVs might return no records!

One starting point is to use the sys.dm_db_index_physical_stats DMV to look at the indexes' size and fragmentation information for a given table (or view). The table is used usually as starting point for analyzing indexes' fragmentation and then defragment the indexes with high fragmentation.

-- sys metadata - index & data size and fragmentation information for the data and indexes of the specified table or view
SELECT --db_name() db_name
--, object_name(IND.object_id) table_name
 IND.name index_name
, IND.type_desc
, IPS.page_count
, IPS.record_count
, IPS.index_level
, Cast(IPS.avg_fragmentation_in_percent as decimal(10,2)) avg_fragmentation_perc
, Cast(IPS.avg_page_space_used_in_percent as decimal(10,2)) space_used_perc
--, IPS.*
FROM sys.indexes IND
     CROSS APPLY sys.dm_db_index_physical_stats(DB_ID(), IND.object_id, IND.index_id, NULL, 'DETAILED') IPS
WHERE IND.object_id = OBJECT_ID(N'SalesLT.Product');

Output:

index_name	type_desc	page_count	record_count	index_level	avg_fragmentation_perc	space_used_perc
PK_Product_ProductID	CLUSTERED	101	295	0	0.99	87.90
PK_Product_ProductID	CLUSTERED	1	101	1	0.00	16.20
AK_Product_rowguid	NONCLUSTERED	2	295	0	50.00	74.69
AK_Product_rowguid	NONCLUSTERED	1	2	1	0.00	0.59
AK_Product_ProductNumber	NONCLUSTERED	2	295	0	50.00	85.79
AK_Product_ProductNumber	NONCLUSTERED	1	2	1	0.00	0.49
AK_Product_Name	NONCLUSTERED	3	295	0	33.33	87.32
AK_Product_Name	NONCLUSTERED	1	3	1	0.00	1.67
IX_SalesLT_Product_Color	NONCLUSTERED	1	295	0	0.00	79.24
IX_SalesLT_Product_Color_Size	NONCLUSTERED	1	295	0	0.00	94.12
IX_SalesLT_Product_ListPrice_IC	NONCLUSTERED	4	295	0	0.00	86.60
IX_SalesLT_Product_ListPrice_IC	NONCLUSTERED	1	4	1	0.00	1.01

In a second step one can look at the sys.dm_db_index_usage_stats DMV which provides the counts of the different types of index operations and the time each type of operation was last performed:

-- sys metadata - counts of different types of index operations and the time each type of operation was last performed.
SELECT -- db_name() db_name
--, object_name(IND.object_id) table_name
 IND.name
, IND.type_desc
, IUS.user_seeks 
, IUS.user_scans
, IUS.user_lookups 
, IUS.user_updates
, IUS.last_user_seek
, IUS.last_user_scan 
, IUS.last_user_lookup
, IUS.last_user_update
FROM sys.dm_db_index_usage_stats IUS
     JOIN sys.indexes IND
       ON IUS.index_id = IND.index_id
WHERE IND.object_id = OBJECT_ID(N'SalesLT.Product');

Output:

name	type_desc	user_seeks	user_scans	user_lookups	user_updates	last_user_seek	last_user_scan	last_user_lookup
PK_Product_ProductID	CLUSTERED	0	10	15	0		2025-01-06T14:23:54	2025-01-06T14:23:54
IX_SalesLT_Product_Color_Size	NONCLUSTERED	11	0	0	0	2025-01-06T14:23:54
IX_SalesLT_Product_ListPrice_IC	NONCLUSTERED	8	0	0	0	2025-01-06T13:38:03

Finally, it might be useful to look also at the sys.dm_db_index_operational_stats DMV which returns the current lower-level I/O, locking, latching, and access method activity for each partition of a table or index in the database (see the documentation for the full list of attrbutes):

-- sys metadata - index operations stats
SELECT -- db_name() db_name
--, object_name(IND.object_id) table_name
 IND.name index_name
, IND.type_desc
, IOS.range_scan_count
, IOS.singleton_lookup_count
, IOS.leaf_insert_count
, IOS.leaf_delete_count
, IOS.leaf_update_count
, IOS.nonleaf_insert_count
, IOS.nonleaf_delete_count
, IOS.nonleaf_update_count
FROM sys.indexes IND
     CROSS APPLY sys.dm_db_index_operational_stats(DB_ID(), IND.object_id, IND.index_id, NULL) IOS
WHERE IND.object_id = OBJECT_ID(N'SalesLT.Product')
 AND IOS.range_scan_count<>0
ORDER BY IND.name;

Output:

index_name	type_desc	range_scan_count	singleton_lookup_count	leaf_insert_count	leaf_delete_count	leaf_update_count	nonleaf_insert_count	nonleaf_delete_count	nonleaf_update_count
IX_SalesLT_Product_Color_Size	NONCLUSTERED	11	0	0	0	0	0	0	0
IX_SalesLT_Product_ListPrice_IC	NONCLUSTERED	8	0	0	0	0	0	0	0
PK_Product_ProductID	CLUSTERED	10	64	0	0	0	0	0	0

For more information on these DMVs check the documentation.

Happy coding!

Previous Post <<||>> Next Post

References:
[1] Microsoft Learn (2024) SQL Server: sys.dm_db_index_physical_stats [link]
[2] Microsoft Learn (2024) SQL Server: sys.dm_db_index_usage_stats [link]
[3] Microsoft Learn (2024) SQL Server: sys.dm_db_index_operational_stats [link]

💎🏭SQL Reloaded: Microsoft Fabric's SQL Databases (Part V: Manual Index Maintenance) [new feature]

Indexes' maintenance in Microsoft Fabric's SQL databases is supposed to happen automatically in the background via automatic tuning options feature, though the whole functionality is still in its early phases, and therefore many questions regarding the whole process may arise. Probably the most important question is whether indexes can still be created, respectively maintained manually. That's useful for temporary or even periodic workloads, where maybe organizations might still want to maintain indexes manually. 

The tests made below are based on the SalesLT.Product from AdventureWorkds database available in Microsoft Fabric. The target was to create several indexes that could be used for the various testing purposes. Each set of the below scripts was run 5-10 times until records appeared in the sys.dm_db_missing_index_details table for each test case (see further below):

-- batch 1: filter on single column (to be run 5-10 times)
SELECT *
FROM SalesLT.Product 
WHERE Color = 'Red'

SELECT *
FROM SalesLT.Product 
WHERE Color = 'Black'

SELECT *
FROM SalesLT.Product 
WHERE Color = 'White'

-- batch 2: filter on two columns (to be run 5-10 times)
SELECT *
FROM SalesLT.Product 
WHERE Color = 'Red'
  AND Size = '58'

SELECT *
FROM SalesLT.Product 
WHERE Color = 'Black'
  AND Size = '58'

SELECT *
FROM SalesLT.Product 
WHERE Color = 'White'
     AND Size = '58'

-- batch 3: filter with column selection (to be run 5-10 times)
SELECT ProductNumber, Name, Color, ListPrice
FROM SalesLT.Product 
WHERE ListPrice BETWEEN 50 AND 55

SELECT ProductNumber, Name, Color, ListPrice
FROM SalesLT.Product 
WHERE ListPrice BETWEEN 100 and 105

Once the scripts run, one can look at the records created in the above considered dynamic management view:

-- sys metadata -  missing indexes
SELECT MID.statement AS table_name
, MID.equality_columns
, MID.inequality_columns
, MID.included_columns
--, MIG.index_group_handle
--, MIG.index_handle
FROM sys.dm_db_missing_index_details MID 
    JOIN sys.dm_db_missing_index_groups MIG 
     ON MID.index_handle =  MIG.index_handle
ORDER BY MIG.index_group_handle
, MIG.index_handle

Output:

table_name	equality_columns	inequality_columns	included_columns
[AdventureWorks01-...].[SalesLT].[Product]	[Color]
[AdventureWorks01-...].[SalesLT].[Product]	[Color], [Size]
[AdventureWorks01-...].[SalesLT].[Product]		[ListPrice]	[Name], [ProductNumber], [Color]

The next step is to create one of the indexes (please note that database's name must be replaced accordingly or used only the 2-part naming convention - schema & table name ):

-- create index on Color
CREATE INDEX IX_SalesLT_Product_Color 
ON [AdventureWorks01-...].[SalesLT].[Product] (Color);

Once the script was run, all the records related to the SalesLT.Product disappeared from the dynamic management view. Therefore, it might be a good idea to take a snapshot with view's data before creating any indexes manually. Probably the same behavior should be expected when the indexes are created by the system.

-- create index on Color & Size
CREATE INDEX IX_SalesLT_Product_Color_Size
ON [SalesLT].[Product] (Color, Size);

-- create index on ListPrice with included columns
CREATE INDEX IX_SalesLT_Product_ListPrice_IC
ON [SalesLT].[Product] (ListPrice) INCLUDE(ProductNumber, Name, Color);

One can use the following query based on the meta.vIndexes (created in a previous post) to look at the indexes created:

-- sys metadata - index columns
SELECt IND.db_name
, IND.schema_name
, IND.table_name
, IND.index_name
, IND.index_type
, IND.principal_type
, IND.auto_created
FROM meta.vIndexes IND
WHERE IND.schema_name = 'SalesLT'
  AND IND.table_name = 'Product'
  AND IND.index_name IN ('IX_SalesLT_Product_Color ','IX_SalesLT_Product_Color_Size'
,'IX_SalesLT_Product_ListPrice_IC')
ORDER BY IND.table_name
, IND.index_name

Output:

db_name	schema_name	table_name	index_name	index_type	principal_type	auto_created
AdventureWorks01-...	SalesLT	Product	IX_SalesLT_Product_Color	NONCLUSTERED	S	False
AdventureWorks01-...	SalesLT	Product	IX_SalesLT_Product_Color_Size	NONCLUSTERED	S	False
AdventureWorks01-...	SalesLT	Product	IX_SalesLT_Product_ListPrice_IC	NONCLUSTERED	S	False

After this model can be created further indexes as needed. It's always a good idea to take a "copy" of the indexes created (or keep a history of the scripts run for indexes' maintenance). This best practice is now more important, when the system can drop indexes as it considers fit. 

Don't forget to clean up the changes made if the indexes aren't needed anymore:

-- cleaning after
DROP INDEX IF EXISTS SalesLT.IX_SalesLT_Product_Color;
DROP INDEX IF EXISTS SalesLT.IX_SalesLT_Product_Color_Size;
DROP INDEX IF EXISTS SalesLT.IX_SalesLT_Product_ListPrice_IC;

So, after these tests, the standard syntax for index's maintenance seems to work also on SQL databases, with all the implications deriving from this (e.g. porting of scripts, database objects, etc.)

Happy coding!

Previous Post <<||>> Next Post

🏭 💠Data Warehousing: Microsoft Fabric (Part VI: SQL Databases for OLTP scenarios) [new feature]

Data Warehousing Series

One interesting announcements at Ignite is the availability in public preview of SQL databases in Microsoft Fabric, "a versatile and developer-friendly transactional database built on the foundation of Azure SQL database". With this Fabric can address besides OLAP also OLTP scenarios, evolving thus from analytics to a data platform [1]. According to the announcement, besides the AI-optimized architectural aspects, the feature makes the SQL Azure simple, autonomous and secure by design [1], and these latest aspects are considered in this post. 

Simplicity revolves around the deployment and configuration of databases, the creation of a new database requiring giving a name and the database is created in seconds [1]. It’s a considerable improvement compared with the relatively complex setup needed for on-premise configurations, though sometimes more flexibility in configuration is needed upfront or over database’s lifetime. To get a database ready for testing one can import a sample database or get specific data via data flows and/or pipelines [1]. As development tools one can use Visual Studio Code or SSMS [1], and probably more tools will be available in time.

The integration with both GitHub and Azure DevOps allows to configure each database under source control, which is needed for many scenarios especially when multiple resources make changes to the database objects [1]. Frankly, that’s mainly important during the development phase, respectively in scenarios in which multiple people make in parallel changes to the logic. It will be interesting to see how much overhead or challenges the feature adds to development and how smoothly everything works together!

The most important aspect for many solutions is the replication of data in near-real time to the (open-source) delta parquet format in OneLake and thus making the data available for analytics almost immediately [1]. Probably, from this aspect many cloud-based applications can benefit, even if the performance might not be as good as in other well-established architectures. However, there are many other scenarios in which one needs to maintain and use data for OLTP/OLAP purposes. This invites adequate testing and a good weighting of the advantages and disadvantages involved. 

A SQL database is a native item in Fabric, and therefore it utilizes Fabric capacity units like other Fabric workloads [1]. One can use the Fabric SKU estimator (still in private preview) to estimate the costs [2], though it will be interesting to see how cost-effective the solutions are. Probably, especially when the infrastructure is already available outside of Fabric, it will be easier and cost-effective to use the mirroring functionality. One should test and have a better estimator before moving blindly from the existing infrastructure to Fabric. 

SQL databases in Fabric are autonomous by design, while allowing to get the best performance and availability by default [1]. High availability is reached through zone redundancy, while performance is achieved by scaling automatically the storage and compute to accommodate the workloads [1]. The auto-optimization capability is achieved with the help of the latest Intelligent Query Processing (IQP) enhancements, respectively the creation of missing indexes to improve query performance [1]. It will be interesting to see how the whole process works, given that the maintenance of indexes usually involves some challenges (e.g. identifying covering indexes, indexes needed only for temporary workloads, duplicated indexes).

SQL databases in Fabric are automatically configured for high availability with zone redundancy, while storage and compute scale automatically to accommodate the user workload [1]. The database is auto-optimized through the latest IQP enhancements while the system creates any missing indexes to improve query performance. All data is replicated to OneLake by default [1]. Finally, the database always receives the latest security updates with auto-patching, while automatic backups help in disaster recovery scenarios  [1], which can be of real help for database administrators. 

References:
[1] Microsoft Fabric Updates Blog (2024) Announcing SQL database in Microsoft Fabric Public Preview [link]
[2] Microsoft Fabric Updates Blog (2024) Announcing New Recruitment for the Private Preview of Microsoft Fabric SKU Estimator [link]