Columnstore Index

The Baker’s Dozen
Business Intelligence
13 Tips for the SQL Server Columnstore Index
Kevin S. Goff
Microsoft SQL Server MVP
Kevin S. Goff – Brief BIO
• Developer/architect since 1987 / Microsoft SQL Server MVP
• Columnist for CoDe Magazine since 2004,
“The Baker’s Dozen” Productivity Series”, 13 tips on a SQL/BI topic
• Wrote a book, collaborated on a 2nd book
• Frequent speaker for SQL Server community events and SQL Live!360
• Email: [email protected]
• My site/blog: www.KevinSGoff.Net (includes SQL/BI webcasts)
• Releasing some SQL/BI video courseware in 2015
Intro to Power BI for Office 365
Columnstore Index - Introduction
• Today: 13 topics for the Columnstore index
• New index in SQL Server 2012, enhanced in SQL 2014
• More than just an index, an in-memory compressed
• A real game-changer, one of the biggest features in the SQL
database engine of all time
• Some companies upgraded to SQL 2012 just because of this
• Represents another example where MS is devoting serious
attention to the underlying database engine
Earlier versions of SQL Server (2005) focused largely on language and developer
SQL 2008 and 2012 have seen underlying database management/engine changes (Change
Data Capture and Columnstore Index)
Columnstore Index - Introduction
• In SQL 2012, not everyone benefits from this
• Built for more Data warehouse/data mart environments , and
even then, only certain ones (in 2012 it’s a READONLY index,
but that changes in 2014)
• For Data Warehousing environments, the columnstore index is
one more reason why Data Warehouses/Data Marts should
shape data in star-schema Fact/Dimension Models with
surrogate integer keys
Columnstore Index - Topics
Quick demonstration
Overview to the Columnstore Index
Characteristics of the Columnstore Index
Who benefits from this?
Columnstore indexes vs Rowstore Index
Execution plan using the Columnstore index
Batch Mode Processing – new processing mode for the Columnstore
8. Where the Columnstore index can’t directly be used
9. Selective vs non-Selective queries – where Columnstore index isn’t used
10.General Usage rules
11.Restriction rules on the Columnstore index
12.Overall Performance Benchmarks
13.New features in SQL Server 2014
1 – Quick Demonstration
Demo code…
2 – Introduction to Columnstore Index
• New relational, xVelocity memory-optimized database index in SQL
Server 2012, “baked in” to the database engine
• xVelocity used to be called VertiPaq, found in PowerPivot going back to
• More and more functionality in DB engine (xVelocity, CDC)
• Potentially Significant performance enhancements for data warehousing
and data mart scenarios – a real game changer
– (not really for OLTP databases, we’ll see why later)
• Best for queries that scan/aggregate large sets of data
• My opinion? One of the coolest things ever in SQL Server
• In a regular index, indexed data from each row kept together on single
page – and the data in each column spread across all pages of index
• In a columnstore index, data from each column is kept together (pages
stored adjacently) so each data page contains data only from a single
column (compressed, more fits in memory, more efficient IO)
3 – Characteristics of Columnstore Index
• Highly compressed - Exploits similarity of data within column
– Typical in data warehouse Fact Table foreign keys
• IO Statistics - dramatically reduces # of logical reads!!!
• Not stored in standard buffer pools, but rather in a new optimized
buffer pool cache and a new memory broker
• Smart IO and caching using aggressive read-ahead read strategy
• Part of Microsoft’s xVelocity technology – compression is factor of
8 (and twice as efficient as page compression)
• Once posted, only READONLY (this changes in SQL 2014)
• Best for data warehouse/mart queries that scan/aggregate large
amounts of data–might lower need for OLAP aggregation
• Some queries might run at least 10x faster (or more)
4 – Who Benefits?
• Queries and reports against Data Warehouses/Data Marts (works best
with Fact/Dimension tables modeled in a star schema)
• Load from Data Warehouses/Marts into OLAP Cubes (more so in SQL
• SSAS OLAP Databases that use the ROLAP methodology or pass-through
mode “might” benefit (more so in SQL 2014)
• New Analysis Services Tabular Model uses xVelocity engine
• Some companies took the release candidate and put into production,
simply for this feature (some case studies show queries that went from
17 minutes to 3 seconds!)
• If you want to see Memory Usage, good blog entry from Joe D’Antoni
(website is SQL Herald, He and other
developers wrote procedure to return amount of memory used by
columnstore object pool
5 – Columnstore vs Rowstore Index
• Columnstore index stores each column in
separate set of pages (vs. storing multiple data
rows per page using b-trees, key values)
• Only columns needed are fetched
• Easier to compress redundant column data
• Uses xVelocity found in PowerPivot
• Improved IO scan/buffer hit rates
• Segment elimination: each partition is broken
into million row segments with metadata for
min/max values – segment is not read if query
scope does not include min/max values
• Query will only fetch necessary columns
• In reality, not “really” an index – more like a
compressed “cube”
5 – Columnstore vs Rowstore Index
Stored as a
vector (value
that determines
position of one
point in space
relative to
• Because a Fact table might contain millions of
rows for a single CustomerFK or ProductFK, SQL
Server can compress all the repeated surrogate
keys to a single value
• Under the hood, SQL Server is not storing the
values of 2, 3, etc….it is storing a special vector an offset value with respect to the prior value (for
• SQL Server also uses segment elimination for rows
not needed – so any query for year of 2011 can
eliminate the segments for 2010 and 2012
• Bottom line: ALL SORTS of efficiency baked into
the engine – but there’s even more!
• This is one more reason to shape data
warehouses/marts into star-schema, FactDimension models with surrogate keys
6 – Execution Plan with Columnstore
• Go back to Execution Plan
• Columnstore index was 5% of the
• Clustered index was 65% of the
• Covering index (which would have
been the best approach prior to
SQL Server 2012) was 35% of the
• Time Statistics, 12x faster than
covering index, 20x faster than
clustered index
7 – Batch Mode Processing
Packets of about 1,000 rows are
passed between operators, with
column data represented as a
“Vector-oriented processing”
Huge reduction in CPU Usage, at
least by a factor of 10 or more
Batch mode takes advantage of
advanced hardware
architectures, processor cache,
and RAM , improves parallelism
Go back to Execution Plan
New Processing Model in
SQL 2012
Certain execution plan
operators (Hash Join and
Hash Aggregation in
particular) use new Batch
execution mode
Reads rows in blocks of
1,000 in parallel, minimizes
instructions per row
Data moves in batches
through query plan
Big performance benefit
over row-based execution
8 – Where Columnstore can’t be used
Demo code…
Issue w/OUTER JOIN: can’t use directly against table
Will “work”, but will use slower row execution mode
Must pre-aggregate separately and then do OUTER JOIN (will use batch mode)
9 – Selective vs Non-Selective queries
Demo code…
10 – General Usage Syntax and Rules
Syntax is simple: use new COLUMNSTORE keyword
1 Columnstore index per table: cannot be clustered (in 2014 can be clustered)
Order of columns does not matter
Include all columns from table
No INCLUDE statement, No ASC/DESC
General MS recommendation: if queries will frequently use a certainly column on
the predicate, create a clustered index on that column and then create the
columnstore index.
Even though column store index isn’t “ordered” itself, you’ll get better segment elimination
ON [BigPurchaseOrderHeader]
(PurchaseOrderID, VendorID, OrderDate,
ShipMethodID, Freight, TotalDue)
• Include all columns
• Order doesn’t
• No key columns
11 – Restrictions and Rules
• Cannot be clustered, cannot be created against a view
• Cannot act as a PK or FK, cannot include sparse columns
• Can’t work on tables with Change Data Capture/Change Tracking or
FileStream, can’t participate in replication, nor when page/row
compression exists
• Cannot be used with certain data types, such as binary, text/image,
rowversion/timestamp, CLR data types (hierarchyID/spatial), nor
with data types created with MAX keyword…e.g. varchar(max)
• Cannot be used with UniqueIdentifier
• Cannot be used with decimal > 18
• Cannot be modified with ALTER – must be dropped and recreated
• It’s a read-only index - cannot insert rows and expect columnstore
index to be maintained (changes in SQL 2014)
11 – Restrictions and Rules (continued)
• Note: range partitioning is supported….(use partitioning to load a
table, index it with a columnstore index, and switch it in as newest
– Partition by day, split the last partition
– Load data into staging table and then create columnstore
– Switch it in (URL reference at end of slides for an example)
– SQL Server 2012 permits 15,000 partitions per table
• Not optimized for certain statements (OUTER JOIN, UNION, NOT IN
• Not optimized for certain scenarios (high selectivity, queries lacking
any aggregations)
• Not optimized for a JOIN statement on a composite set of columns
(truthfully, a join between a fact table and a dimension table should
only be on one integer key)
• Best practice – always use integer keys for FKs
12 – General Benchmarks
CPU time
Total Time
Logical Reads
Clustered index
Non-clustered covering index
Column Store index
Column Store
Total Time (ms)
Logical Reads
13 – Enhancements in SQL 2014
New Clustered Columnstore Index (CCI)
• No columns specified - The CCI “is” the data
– It’s Updateable, No longer a read-only index
– Cannot have any non-clustered indexes
– Cannot have key constraints
So you have one of two options:
– A non-clustered read-only columnstore index, index plus as many non-clustered
indexes for FK values as you need (2012 model)
– A clustered read-write columnstore index, but no non-clustered indexes for
specific FK values (2014 model)
• 2014 CCI Index might perform better against highly selective queries than 2012
columnstore indexes did on highly selective queries
Support for more data types
Basically all data types except CLR, varchar(max) and varbinary(max), XML, and spatial
data types
Additional Archive compression on top of regular columnstore compression
Some ask – what is difference between this and Hekaton in-memory optimized tables?
Recommended Links
A great video on Columnstore Index from Tech-Ed 2013:
Adding data to a table using Partition Switching
Last year I did a 13-part series on new features in SQL 2012 for TechNet:
I’ve written some articles in CoDe Magazine on SQL 2012
2 part series on Columnstore index, T-SQL Features, and SSIS Features

similar documents