Azure SQL Data Warehouse

1982 I started working with computers
1988 I started my professional career in computers industry
1996 I started working with SQL Server 6.0
1998 I earned my first certification at Microsoft as
Microsoft Certified Solution Developer (3rd in Greece)
1999 I started my career as Microsoft Certified Trainer (MCT) with
more than 30.000 hours of training until now!
2010 I became for first time Microsoft MVP on Data Platform
I created the SQL School Greece www.sqlschool.gr
2012 I became MCT Regional Lead by Microsoft Learning Program.
2013 I was certified as MCSE : Data Platform
I was certified as MCSE : Business Intelligence
2016 I was certified as MCSE: Data Management & Analytics
Antonios
Chatzipavlis
SQL Server Expert and Evangelist
Data Platform MVP
MCT, MCSE, MCITP, MCPD, MCSD, MCDBA,
MCSA, MCTS, MCAD, MCP, OCA, ITIL-F

Μια πηγή ενημέρωσης για τον Microsoft SQL Server
προς τους Έλληνες IT Professionals, DBAs,
Developers, Information Workers αλλά και απλούς
χομπίστες που απλά τους αρέσει ο SQL Server.
Help line : help@sqlschool.gr
• Articles about SQL Server
• SQL Server News
• SQL Nights
• Webcasts
• Downloads
• Resources
What we are doing here Follow us in socials
fb/sqlschoolgr
fb/groups/sqlschool
@antoniosch
@sqlschool
yt/c/SqlschoolGr
SQL School Greece group
S E L E C T K N O W L E D G E F R O M S Q L S E R V E R

Azure SQL Data Warehouse SQLschool.gr GWAB Athens 2017
Presentation Content
5
• First Look on Azure SQL DW
• Designing for Azure SQL DW
• Loading Data on Azure SQL DW
• Querying and Tuning Azure SQL DW

First Look on
Azure SQL Data Warehouse
6

What is Azure SQL Data Warehouse?
7
Service in
Microsoft Azure
It’s a PAAS
offering
It’s a Massively
Parallel Processing
System
Distribute
Storage
Distributed
Compute

SMP vs MPP
8
Symmetric Multiprocessing Massively Parallel Processing

Data Warehousing Unit
9
A measure of the
underlying compute
power of the database

Data Warehousing Unit
10
For Example
50 100
100 DWU 500 DWU
3 table loaded in 15 min
20 minutes to run a report
3 table loaded in 3 min
4 minutes to run a report

Why Choose Cloud Over On-Premises DW?
11
• Doesn’t need large CAPEX to get started
• Doesn’t need large OPEX
• We can scale storage and compute up or down
on demand

What and How do you pay for this Service ?
12
• Storage
– Storage is billed by GB
– Standard or Premium Geo Redundant
– No cost for storage transactions
– Outbound data transfer is billed
• Compute Power
– Compute is billed by DWUs
– Can go from 100 to 2000
– Billed per hour
When not in use, compute
power of the DW can be
completely paused for
maximum savings

Provisioning Azure SQL Data Warehouse
13
Select a
Region
Select or
Create a
Server
Pick
origin of
the data
Pick
DWU
level

Methods of Provisioning
14
• Azure Portal
– Select New > Data + Storage
• PowerShell
– New AzureRmSqlDatabase Cmdlet
• T-SQL
– CREATE DATABASE Command

Provision a Data Warehouse
15
DEMO

Designing for
16

SQL Server Azure SQL DW!=
An Azure SQL DW database requires design
decisions that are different from SQL Server

Distribution Key
18
Determines the method in which Azure
SQL Data Warehouse spreads the data
across multiple nodes
uses up to 60 distributions
when loading data into the
system

Hash Distribution
19
RecordNo CustomerID InvoiceDate
1 1000 2017-04-21
2 1000 2017-04-22
3 2000 2017-04-22
4 3000 2017-04-22
5 4000 2017-04-22
Hashing by CustomerID

Round-Robin Distribution
20
RecordNo CustomerID InvoiceDate
1 1000 2017-04-21
2 1000 2017-04-22
3 2000 2017-04-22
4 3000 2017-04-22
5 4000 2017-04-22
Rows distributed to all nodes

Data Distribution best practice
21
Even DistributionOdd Distribution

Good Hash Key
22
Distributes
Evenly
Used for
Grouping
Used as
Join Condition
Is Not
Updated
Has more than
60
distinct values
Round-Robin will always provide a uniform distribution but not necessarily the best performance

Data Types
23
Use the smallest
data type which will
support your data
Avoid defining all
character columns
to a large default
length
Define columns as
VARCHAR instead of
NVARCHAR if you
don’t need Unicode
The goal is to not only save space but also move data as efficiently as possible
Some complex data types (xml, geography, etc) are not supported yet

Table Types
24
Clustered
Columnstore
Default table
type
High
compression
ratio
Ideally
Segments of
1M rows
No secondary
indexes Heap
No index on
the data
Fast Load
No
compression
Allows
secondrary
indexes
Clustered
B-Tree
Sorted index
on the data
Fast singleton
lookup
No
compression
Allows
secondary
indexes

Table Partitioning
25
1. Ease of loading and removal of data from a partitioned table
2. Targeting specific partitions on table maintenance operation
3. Performance improvements due to partition elimination
Partitioning is very common in SQL Server Data Warehouses for three reasons:
A highly granular partitioning scheme can work in SQL Server but hurt performance in Azure SQL DW
60 Distributions 365 Partitions 21.900 Data Buckets
21.900 Data Buckets Ideal Segment Size
(1M Rows)
21.900.000.000
Rows
Lower Granularity (week, month) can perform better depending on how much data you have

How do we apply these principles to a Dimensional Model?
26
• Fact Table
– Large ones are better as Columnstores
– Distributed through Has key as much as possible as long as it is even
– Partitioned only if the is large enough to fill up each segment
• Dimension Tables
– Can be Hash distributed or Round-Robin if there is no clear candidate
join key
– Columnstore for large dimensions
– Heap or Clustered Index for small dimensions
– Add secondary indexes for alternate join columns
– Partitioning not recommended

Analyzing distribution and data types for DW tables
27
DEMO

Loading Data on
28

Loading an MPP System
29
The main principle of loading
data into Azure DW is to do as
much work in parallel as possible

Data Warehouse Readers
30
100 200 300 400 500 600 1000 1200 1500 2000
Readers 8 16 24 32 40 48 60 60 60 60
Writers 60 60 60 60 60 60 60 60 60 60
DWU
Your DWUs have a direct impact on how fast you can load data in parallel
- Azure SQL Data Warehouse introduces the concept of Data Warehouse
Readers.
- These are threads that will be reading data in parallel and then passing it
off to Writer threads.

Optimize Insert Batch Size
31
• Avoid trickle insert pattern
– Ideal batch size is 1 million or more direct or in a file
• Avoid Ordered Data
– Data ordered by distribution key can introduce hot spots that slow down the load
operation
• Using Temporary Tables
– Stage and transform on a Temp Heap table before moving to permanent storage
• Use the CREATE TABLE AS statement
– Fully parallel operation
– It’s minimally logged
– It can change: distribution, table type, partitioning

CREATE TABLE #fact_tmp
WITH
(
DISTRIBUTION = ROUND_ROBIN
)
AS
SELECT *
FROM dbo.FactInternetSales;

User Resource Class
33
Class Smallrc Mediumrc Largerc Xlargerc
Default 8 16 24 32
Memory 100 MB 100-1600 MB 200-3200 MB 400-6400 MB
The lower range corresponds to DWU100 the upper range to DWU2000
User Resource classes as database roles that govern how many resources
are given to a query
For fast and high quality loads create a user just for loading which utilize a medium or large RC

Loading Methods
34
• Single-client loading methods
– SSIS
– Azure Data Factory
– BCP
– Can add some parallel capabilities but are bottleneck at the Control node
• Parallel readers loading methods
– PolyBase
– Reads from Azure Blob Storage and loads the content into Azure SQL DW
– Bypasses the Control node and loads directly into the Compute Nodes

Control Node
35
The Control Node
receives connections
and orchestrates the
queries
The Compute Nodes
do processing on the
data and scale with
the DWUs

Loading with SSIS
36
SSIS Control
Node

Loading data with SSIS
37
DEMO

Loading with PolyBase
38
Control
Node Azure
Blob Storage
PolyBase can load data from
UTF-8 delimited text files and
popular Hadoop file formats
(RC file, ORC and Parquet)

Loading data with PolyBase
39
DEMO

Migration Utility
40
• Supports SQL Server 2012+ and
Azure SQL Database
• Provides a migration report pointing
out possible issues
• Assists with schema migration
• Assists with data migration

Using the Azure SQL DW migration utility
41
DEMO

Querying and Tuning
42

Workload Management Principles
43
User Resource
Class
Concurrency
Model
Transaction Size
TwoMaximumLimits
1024 Connections
32 Concurrent Queries

Concurrency Queries and Concurrency Slots
44
100 200 300 400 500 600 1000 1200 1500 2000
Slots 4 8 12 16 20 24 40 48 60 80
DWU
Queries Executing Queries Incoming Queries Queued
DW200 7 2 1
DW1000 32 2 2
Examples
The above examples assumes that each query is consuming 1 concurrency slot

Resource Class and Concurrency Slots
45
Class Smallrc Mediumrc Largerc Xlargerc
DWUs 100-2000 100-2000 100-2000 100-2000
Slots 1 1-6 2-32 4-64
SELECT queries against system views, stats and other management commands do not use concurrency slots

Transaction Size Limits
46
100 200 300 400 500 600 1000 1200 1500 2000
GB /
Distribution
1 1,5 2,25 3 3,75 4,5 7,5 9 11,25 15
DWU
A DW200 transaction doing equal work per distribution could
consume 60 x 1,5 GB = 90 GB of space

Maintaining Statistics
47
• The service does not create or maintain stats
automatically 
• Creating New stats
– Sampled single column stats is a good start
– Multi columns stats for joins involving multiple columns
– Focus on columns used in JOINs, GROUP BY, HAVING and WHERE clauses
– Increase the sample if necessary
• Updating existing stats
– If new dates or dimension categories added
– If new data loads have completed
– If an UPDATE or DELETE changes the distribution of data

Index Defrag
48
• Heap
– Does not have a defrag option
• B-Tree Index
– Useful for removing low levels of fragmentation
• Columnstore
– Proactively compresses CLOSED rowgroups
• On a large table with heavy fragmentation it is often faster to recreate the
table with the CREATE TABLE AS SELECT and switch it with the older

Index Rebuild
49
• Heap
– Can be rebuilt to remove forward pointers
• B-Tree Index
– Will remove high levels of fragmentation
• Columnstore
– Can increase the density of segments
• Rebuilding as index is an OFFLINE operation in Azure SQL DW

Scaling Performance
50
• Increase the User Resource Class
– EXEC sp_addrolemember ‘largerc’, ‘loaduser’;
– Higher Resource Class – more memory and CPU
– More concurrency slots – less concurrent queries
– The highest role assigned takes precedence
• Increase the Data Warehouse Units
– ALTER DATABASE AWDW MODIFY (SERVICE_OBJECTIVE=‘DW1000’);
– It is an OFFLINE operation
– Make sure there are no loads or transactions in progress
– Can also be done through the Azure Portal

Tracking Queries with Lables
51
SELECT sum(Qty)
FROM dbo.FactInternetSales
OPTION (LABEL=‘mylabel’);
SELECT *
FROM sys.dm_pdw_exec_requests
WHERE label=‘mylabel’);
User Query
Admin Query

Labeling a query and tracking its execution
52
DEMO

Thank You
S E L E C T K N O W L E D G E F R O M S Q L S E R V E R

Azure SQL Data Warehouse

More Related Content

What's hot

What's hot (20)

Viewers also liked

Viewers also liked (10)

Similar to Azure SQL Data Warehouse

Similar to Azure SQL Data Warehouse (20)

More from Antonios Chatzipavlis

More from Antonios Chatzipavlis (20)

Recently uploaded

Recently uploaded (20)

Azure SQL Data Warehouse