1. Before we begin Goals Non-goals (but feel free to ask questions)
9/17/ :59 PM
Before we begin
Goals
Columnstore fundamentals
Health and monitoring
Performance and scalability
Loading patterns and tools
Non-goals (but feel free to ask questions)
Every detail about SQL Data Warehouse and columnstore
Columnstore deep dive (deeper than design and behavior)
Read every bullet in every slide
Pre-requisites
Working knowledge of SQL Server and SQL Data Warehouse
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

2. Getting peak performance from your SQL Data Warehouse columnstore
9/17/ :59 PM
BRK4016
Getting peak performance from your SQL Data Warehouse columnstore
Shivani Gupta, Joe Yong
SQL Data Warehouse Program Management
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

3. Agenda Azure SQLDW Columnstore Primer Columnstore Health
9/17/ :59 PM
Agenda
Azure SQLDW Columnstore Primer
Terminology & Operations (Insert, Update, Delete, Scan)
Distributions and Partitions
Columnstore Health
Pressures (memory and dictionary)
Monitoring for health
ELT/ETL Patterns and Guidance
Maximizing row group quality
Partitioning guidance, Ordering
Secondary B-Tree Indexes
Data Loading Tools
Polybase, SSIS, Data Factory, BCP
© 2016 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

4. Azure SQL Data Warehouse
9/17/ :59 PM
Full managed, data warehouse as a service
Massively Parallel Processing (MPP) architecture
Separation of storage and compute
Industry leading SQL Server in each compute node
Control
Node
Compute
SQL DB
Blob storage [WASB(S)]
Massively Parallel Processing (MPP) Engine
Azure Infrastructure and Storage
Application or User connection
Data Loading
(PolyBase, ADF, SSIS, REST, ODBC, ADF, AZCopy)
DMS
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

5. SQLDW Columnstore Primer
9/17/ :59 PM
SQLDW Columnstore Primer
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

6. Columnstore Index: Why?
9/17/2018
Columnstore Index: Why?
Data stored as rows
Data stored as columns C5 C1 C2 C3 C4
rowgroup …
Ideal for OLTP
Frequent read/write on small set of rows
Ideal for Data Warehouse
Analytics on large number of rows
segment
Improved compression:
Data from same domain compress better
Reduced I/O:
Fetch only columns needed
Improved Performance:
More data fits in memory
Optimized for CPU utilization
Batch Mode Execution
Vector Processing
© 2015 Microsoft Corporation. All rights reserved. Microsoft, Windows, and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries. The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

7. Terminology Row group Column Segment Dictionary Metadata
9/17/ :59 PM
Terminology
1 million row “row group”
Row group
Set of rows handled together, usually 1M
Column Segment
Values from one column of the row group
Dictionary
Encoding for string columns to integers
Primary dictionary: shared across row segments
Secondary dictionary: local to segment (max 1)
Metadata
Min,max values for segment elimination
dictionaries
primary
secondary
Column Segments
metadata
Min
Max
Col 1 val
Col1 val
Col 2 val …
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

8. 9/17/ :59 PM
Load/Insert
Batch inserts >= 100K rows go directly to compressed row groups
Row groups can be trimmed pre- maturely under memory pressure.
Batch inserts < 100k rows go to delta row group (which is a B- tree).
Delta RGs are closed after 1 M rows.
Tuple Mover moves closed delta RGs to compressed row groups. C1 C2 C3
Delta RGs
Closed
Open
Tuple Mover
Batch Size < 100k
Batch Size >= 100k
Compressed RGs
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

9. DML – delete and update Rows in delta stores are deleted directly
Rows in compressed row groups are not removed
Their locators are tracked in the delete bitmap
Update = delete + insert
Delete bitmap

10. Scan
Scans need to combine data from compressed row groups, delta stores and delete bitmaps to produce correct results
Segment metadata used to eliminate row groups that do not qualify

11. Column Store in SQLDW 60 Distributions
9/17/ :59 PM
Column Store in SQLDW
60 Distributions
Each Distribution has its own Columnstore per table
Multiple row groups per Columnstore
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

12. 9/17/ :59 PM
Partitioning
Each distribution table partitioned N-ways. Total number of Columnstores = 60 * N
CREATE TABLE [cso].[FactOnlineSales_PTN]
( [OnlineSalesKey] int NOT NULL
, [DateKey] datetime NOT NULL
, [StoreKey] int NOT NULL
, [ProductKey] int NOT NULL
, [CurrencyKey] int NOT NULL
, [SalesQuantity] int NOT NULL
, [SalesAmount] money NOT NULL
, [UnitPrice] money NULL )
WITH
( CLUSTERED COLUMNSTORE INDEX
, DISTRIBUTION = HASH([ProductKey])
, PARTITION (
[DateKey] RANGE RIGHT FOR VALUES
' :00:00.000‘
,' :00:00.000' ;
© 2014 Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

13. Partitioning and Column Store in SQLDW
Table is defined with 2- way partitioning
Each distribution’s data split into 2 partitions
2-way partitioning => 120 total Columnstores

14. Columnstore DMVs sys.pdw_nodes_column_store_row_groups
State of various row groups (open, closed, compressed)
Total rows
sys.pdw_nodes_column_store_segments
Encoding used
Dictionary id
Min,max metadata
sys.pdw_nodes_column_store_dictionaries
On disk dictionary size

15. Demo: Primer Concepts

16. Columnstore Health

17. What determines scan performance?
Size of compressed row groups (aka rowgroup quality)
Compression quality decreases with fewer rows
Per row overhead of ancillary structures increases with fewer rows
1M rows per row group ideal
100,000 rows per row group acceptable
< 10,000 abysmal. Disallowed now.
Number of rows in delta stores
Delta store is scanned row by row
Small batch inserts end up in delta stores
Singleton inserts end up in delta stores
Higher DOP increases number of delta stores
Size of Delete Bitmap
Larger delete bitmap implies more time spent merging
Large amount of delete and/or update activity grows delete bitmap

18. Factors affecting Row Group Quality
Available Rows
1. Load Batch Size
2. Partitioning
Premature Trimming
1. Memory Pressure
2. Dictionary Pressure

19. Understanding Memory Pressure
Memory Required for quality row groups
Target rows in the rowgroup (ideal is 1,048,576 = ~1 million rows per rowgroup)
Fixed overhead
#columns in the table
#short string character typed columns (string data types of <=32 bytes)
#long string character typed columns (string data types of >32 bytes)
See details in rowgroup memory
Memory Available to build row groups
SLO of SQL Data Warehouse
User’s Resource Class
Complexity of load query
Partitioning
Degree of Parallelism
Trimming occurs if Required > Available

20. Est. mem grant example #1: 10 column table

21. Memory Management (MB per distribution)
Service admin account

22. Degree of Parallelism (DOP)
Statement memory grant divided b/w parallel workers
Use hint to force serial execution for CTAS (if needed)
Create Table MyFactSalesQuota as Select * from FactSalesQuota OPTION(MAXDOP 1);
< DWU 2000 – DOP 1
DWU – DOP 2
DWU – DOP 3
DWU – DOP 6

23. Dictionary Pressure High cardinality (distinct values) Wide strings
MAX
Dictionary size
16MB in memory
DMV
Dictionary size
Size on disk
On disk size
<
In memory size

24. DMV for Columnstore Health
sys.dm_pdw_nodes_db_column_store_row_group_physical_stats
Row Group State
No of rows
Trim Reason
More details at

25. Demo: Columnstore Health

26. ELT/ETL Patterns and Guidance

27. ELT for maximizing scan performance
Sizing batch loads
Partitioning Guidance
Using Correct Resource Class
Working around Dictionary Pressure
Judicious use of updates and deletes
Secondary B-Tree Indexes (NCI): needle-in-haystack queries
How to improve segment elimination

28. Sizing Batch Loads Target >100,000 per columnstore in each load.
With no partitions, this means > 100,000 * 60 rows (~6 million) rows per CTAS or bulk insert
With 4 partitions , this means > 100,000 * 60 * 4 (~24 million) rows per CTAS or bulk insert
Ideally 1 million per Columnstore in each load
With no partitions, this means > 1,000,000 * 60 rows (~60 million) rows per CTAS or bulk insert
With 4 partitions , this means > 1,000,000 * 60 * 4 (~240 million) rows per CTAS or bulk insert

29. Batching trickle loads
Scenario: 100,000 rows per Columnstore, no partitions
6,000,000+ rows required for each batch load
500
Rows/Sec
1000
2000
Load
threshold exceeded (hours)
<3.5 hours
< 2 hours
< 1 hour

30. Partitioning Guidance
Rowgroups cannot cross partition boundaries
Over-partitioning impacts row group quality
Over-partitioning impacts compression
Make sure your targeted data set allows for at least 6 million rows per partition. Ideally 60 million rows per partition!
Partitioning impacts memory grant

31. Using Correct Resource Class
Compute memory required for quality row groups
Memory guidance for Columnstore
Use views provided.
Grant sufficient memory
Use correct resource class (usually not smallrc)
Scale DWU if needed
Keep load query simple – stage to a Heap/CI if needed
Force serial execution if needed
Don’t Over Allocate Resource Class
Does impact concurrency (especially if multiple ELT jobs scheduled in parallel)

32. Working around Dictionary Pressure
Isolate problematic string columns into a separate table 
Optimize column types where possible (e.g. use varchar(36) instead of nchar(255), smallint instead of nvarchar(4000)

33. Judicious use of Updates and Deletes
Use Heap/CI staging tables for transformations
Alter Index Reorganize/Rebuild to defragment
Reorganize is lighter weight and online
Rebuild is heavy weight and offline

34. Secondary B-Tree Indexes (NCI)
Needle in a haystack queries (equality or short range)
create index l_orderkeynci on lineitem(l_orderkey);
select l_orderkey, l_returnflag, l_linestatus
from lineitem
where l_orderkey = ;
Runs in under a second on 1TB TPC-H data
BTREE (NCI)
Delete bitmap
Delta RGs
Compressed RGs

35. Real world usage of NCI - Yammer
Run a DW 6000
Track all user activity to A/B test every new feature
Fact Events – 300 billion rows (~40TB)
Built NCI on UserId
Query went from 10+ hours to < 5mins:
SELECT TOP 1000 * FROM fact_events WHERE user_id=' '

36. Segment Elimination Min, max metadata used to filter out segments
Columnstore is not ordered
Helps if data arrives naturally ordered e.g. timestamp
Index Rebuild will not keep ordering intact
Min
Max
Col 1 val
Col1 val
Col 2 val …

37. Ordered CCI for improved segment elimination
(Date)
CCI
Partition Switch
ALTER INDEX REBUILD compresses across rowgroups
Removes any manual ordering from CCI

38. Data Loading Tools

39. Data Loading Options PolyBase SSIS* ADF BCP SQLBulkCopy API
Attunity Cloudbeam
ASA/Storm**
Method
Performance
PolyBase
SSIS
ADF
BCP
SQL Bulkcopy
Rate
Rate increases with higher DWU
Yes
Yes* No
Rate increases with more concurrent loads
Fastest Slowest
* With SSIS Azure Feature Pack June 2017 or newer
** Not a good idea

40. Single gated client loading with SSIS data flows (before)
Using SSIS, customers can create data flows with ADO.NET/OLEDB Destination to load data into Azure SQL DW
Similar to loading data into Azure SQL DB/SQL Server
This method has two bottlenecks on the single SSIS-running machine and Control node

41. Single gated client loading with SSIS data flows (before)
Customers can also execute parallel loads with multiple SSIS-running machines by
Divide input data into multiple sources
Load concurrently into separate temporary tables
Switching them in into partitions of the full final table
Overall throughput is still limited by Control node

42. Parallel loading with PolyBase – T-SQL script
Configure the credentials to access your Azure Blob Storage
CREATE DATABASE SCOPED CREDENTIAL AzureStorageCredential WITH
IDENTITY = 'user',
SECRET = '<azure_storage_account_key>';
Define the data source in your Azure Blob Storage with previously configured credentials
CREATE EXTERNAL DATA SOURCE AzureStorage WITH (
TYPE = HADOOP,
LOCATION = CREDENTIAL = AzureStorageCredential);
Define the file format of your input data
CREATE EXTERNAL FILE FORMAT TextFile WITH (
FORMAT_TYPE = DelimitedText,
FORMAT_OPTIONS(FIELD_TERMINATOR = ','));

43. Parallel loading with PolyBase – T-SQL script
Define the external table for your input data
CREATE EXTERNAL TABLE dbo.DimDate2External (
DateId INT NOT NULL,
CalendarQuarter TINYINT NOT NULL,
FiscalQuarter TINYINT NOT NULL)
WITH (
LOCATION='/datedimension/',
DATA_SOURCE=AzureStorage,
FILE_FORMAT=TextFile);
Load the external table from your Azure Blob Storage into Azure SQL DW
CREATE TABLE dbo.DimDate2 WITH (
CLUSTERED COLUMNSTORE INDEX,
DISTRIBUTION = ROUND_ROBIN)
AS SELECT * FROM [dbo].[DimDate2External];

44. Parallel loading with PolyBase (with Azure feature pack)
Create Execute SQL task that triggers PolyBase to load data into Azure SQL DW, leveraging MPP architecture
Throughput is not limited by Control node; scales out based on DWU level

45. Parallel loading with PolyBase
Parallel loading with PolyBase is the recommended method to load large volumes of data into SQL DW
This requires users to manually create/deploy/run multiple tasks on SSIS and/or write T-SQL scripts
Create/deploy/run transformation tasks to convert your input data into supported formats if necessary
Create/deploy/run Azure Blob Upload Task to load it into Azure Blob Storage
Create/deploy/run Execute SQL Task to convert it into external table(s) and trigger PolyBase to load them into Azure SQL DW
Create/deploy/run clean-up tasks if necessary
We have released Azure SQL DW Upload Task to automate most of these tasks (2 – 4)
Leverage existing SSIS expertise
Relieves customers from the “burden” of writing T-SQL script
Azure Blob Storage
SQL DW
Cloud
On-Premise
SQL Server
Flat File
SSIS Machine
1. Export to a flat file
2. Upload to Azure Blob
3. Create external tables and trigger PolyBase to load them
4. Clean up staging area

46. Key takeaways

47. Wrap Up: Key Takeaways
Columnstore behaviour is very different from row store
Loading methods and efficiency matter a lot
Monitor and maintain Columnstore health for better query performance
Assess guidance and ELT patterns for your environment and implement for maximum performance, IF appropriate
Picking the right loading tools can make a BIG difference

48. Please evaluate this session
Tech Ready 15
9/17/2018
Please evaluate this session
From your
Please expand notes window at bottom of slide and read. Then Delete this text box.
PC or tablet: visit MyIgnite
Phone: download and use the Microsoft Ignite mobile app
Your input is important!
© 2012 Microsoft Corporation. All rights reserved. Microsoft, Windows, Windows Vista and other product names are or may be registered trademarks and/or trademarks in the U.S. and/or other countries.
The information herein is for informational purposes only and represents the current view of Microsoft Corporation as of the date of this presentation. Because Microsoft must respond to changing market conditions, it should not be interpreted to be a commitment on the part of Microsoft, and Microsoft cannot guarantee the accuracy of any information provided after the date of this presentation. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.

49. 9/17/ :59 PM
© Microsoft Corporation. All rights reserved. MICROSOFT MAKES NO WARRANTIES, EXPRESS, IMPLIED OR STATUTORY, AS TO THE INFORMATION IN THIS PRESENTATION.