Data types storage differences

It has been a while since I wrote some of my best practices posts. I decided to revisit these posts again to see if anything has changed, I also wanted to see if I could add some additional info.

Today we are going to take a look at how data types can have an impact in queries and also the size of your database.

Char vs NChar

SQL Server has two data types to store character data[1], both of them come in fixed and variable length sizes. The char and varchar data type uses one byte of store to store one character, the nchar and nvarchar data type uses two bytes of store to store one character. The nchar and nvarchar data types are used to store unicode of data

Let's think about that for a second, what we are saying is that the char and varchar data type can store twice the number of characters in the same amount of store as the nchar and nvarchar data type. Why does this matter, space is cheap right? True, space is getting cheaper but we are also storing more and more data every year.

Now think about what happens you have everything stored as unicode data

• What happens to your backup and restore process, will it be faster or slower, will the files be bigger if not compressed?
• What about when transferring the results to and from your database server, are the packets able to store the same number of characters.
• What about the amount of data on a page, what does this do to indexes and index lookups, how does it affect index maintenance?

If you don't need it, then don't use unicode data.
Some examples of what I have seen stored in nchar and nvarchar when realy you shouldn't:

Zip Code for US addresses
US addresses
Social Security Numbers (which were stored in plain text none the less)
Integer data (enforced by constraints or the app layer to make sure these were only digits)

Let's take a quick look by running some T-SQL

First create these two tables

CREATE TABLE TestChar (SomeCol char(10))
GO

CREATE TABLE TestNChar (SomeCol nchar(10))
GO

CREATE index ix_test on TestChar(SomeCol)
GO
CREATE index ix_test on TestNChar(SomeCol)
GO

Now populate each with some data

INSERT TestChar
SELECT TOP 1000000 '1234567890'
FROM sys.sysobjects c1
CROSS JOIN sys.sysobjects c2
CROSS JOIN sys.sysobjects c3
CROSS JOIN sys.sysobjects c4
GO

INSERT TestNChar
SELECT TOP 1000000 '1234567890'
FROM sys.sysobjects c1
CROSS JOIN sys.sysobjects c2
CROSS JOIN sys.sysobjects c3
CROSS JOIN sys.sysobjects c4
GO

Let's see how much space is used by each tables

EXEC sp_spaceused 'TestChar'
EXEC sp_spaceused 'TestNChar'

42768 KB
62736 KB

If you looked at the reserved column, you will see that the nchar data is using about 20 MB more than the char data

Implicit conversions

Besides the storage increase there is also a problem when querying for data that looks like varchar but is stored as unicode. Run the code below.

SET SHOWPLAN_TEXT ON
GO
DECLARE @v varchar(10) = '0123456789'

SELECT * FROM TestChar WHERE SomeCol LIKE  @v +'%'
GO

SET SHOWPLAN_TEXT OFF
GO

Here is the plan for that query

  |--Nested Loops(Inner Join, OUTER REFERENCES:([Expr1008], [Expr1009], [Expr1010]))
       |--Compute Scalar(DEFINE:([Expr1008]=LikeRangeStart([@v]+'%'), [Expr1009]=LikeRangeEnd([@v]+'%'), [Expr1010]=LikeRangeInfo([@v]+'%')))
       |    |--Constant Scan
       |--Index Seek(OBJECT:([Performance].[dbo].[TestChar].[ix_test]), SEEK:([Performance].[dbo].[TestChar].[SomeCol] > [Expr1008] AND [Performance].[dbo].[TestChar].[SomeCol] < [Expr1009]),  WHERE:([Performance].[dbo].[TestChar].[SomeCol] like [@v]+'%') ORDERED FORWARD)

If we look at the plan we can see that this looks pretty good

Usually people will sometimes change the datatype of a column but will not change any code that access this column. Let's now change the data type of the column to nchar

DROP INDEX TestChar.ix_test
GO

ALTER TABLE TestChar ALTER COLUMN SomeCol nchar(10)
GO

CREATE INDEX ix_test on TestChar(SomeCol)
GO

Run the query that gives you the plan again

SET SHOWPLAN_TEXT ON
GO
DECLARE @v varchar(10) = '0123456789'

SELECT * FROM TestChar WHERE SomeCol LIKE  @v +'%'
GO

SET SHOWPLAN_TEXT OFF
GO

Here is the plan

  |--Nested Loops(Inner Join, OUTER REFERENCES:([Expr1008], [Expr1009], [Expr1010]))
       |--Compute Scalar(DEFINE:([Expr1008]=LikeRangeStart(CONVERT_IMPLICIT(nvarchar(11),[@v]+'%',0)), [Expr1009]=LikeRangeEnd(CONVERT_IMPLICIT(nvarchar(11),[@v]+'%',0)), [Expr1010]=LikeRangeInfo(CONVERT_IMPLICIT(nvarchar(11),[@v]+'%',0))))
       |    |--Constant Scan
       |--Index Seek(OBJECT:([Performance].[dbo].[TestChar].[ix_test]), SEEK:([Performance].[dbo].[TestChar].[SomeCol] > [Expr1008] AND [Performance].[dbo].[TestChar].[SomeCol] < [Expr1009]),  WHERE:([Performance].[dbo].[TestChar].[SomeCol] like CONVERT_IMPLICIT(nvarchar(11),[@v]+'%',0)) ORDERED FORWARD)

As you can see, there is a conversion going on right now.

In order to get rid of the conversion, use the correct data types

SET SHOWPLAN_TEXT ON
GO
DECLARE @v nvarchar(10) = '0123456789'

SELECT * FROM TestChar WHERE SomeCol LIKE  @v +'%'
GO

SET SHOWPLAN_TEXT OFF
GO

  |--Nested Loops(Inner Join, OUTER REFERENCES:([Expr1008], [Expr1009], [Expr1010]))
       |--Compute Scalar(DEFINE:([Expr1008]=LikeRangeStart([@v]+N'%'), [Expr1009]=LikeRangeEnd([@v]+N'%'), [Expr1010]=LikeRangeInfo([@v]+N'%')))
       |    |--Constant Scan
       |--Index Seek(OBJECT:([Performance].[dbo].[TestChar].[ix_test]), SEEK:([Performance].[dbo].[TestChar].[SomeCol] > [Expr1008] AND [Performance].[dbo].[TestChar].[SomeCol] < [Expr1009]),  WHERE:([Performance].[dbo].[TestChar].[SomeCol] like [@v]+N'%') ORDERED FORWARD)

Implicit conversions also were an issue when ORMs first burst onto the scene. If you used NHibernate or LINQ to SQL with .NET, since strings in .NET are unicode, all text would be sent over as unicode and you would see all kinds on conversions.

Using larger datatypes when it is not needed

I see this problem mostly with the integer data types. Below is a list of the integer data types together with their storage size and range

tinyint
Storage size is 1 byte. Integer data from 0 through 255.

smallint
Storage size is 2 bytes. Integer data from -2^15 (-32,768) through 2^15 - 1 (32,767).

int
Storage size is 4 bytes. Integer data from -2^31 (-2,147,483,648) through 2^31 - 1 (2,147,483,647).

bigint
Storage size is 8 bytes. Integer data from -2^63 (-9,223,372,036,854,775,808) through 2^63-1 (9,223,372,036,854,775,807).

Now imagine facebook with a billion users decided to use bigint as CountryID in their Country table, this key is then uses as a foreign key in the user demographics table. This is wasteful,either use a smallint since we won't go through 32 thousand countries in the foreseeable feature or use the 2 or 3 character ISO code. 

The problem is even worse if you have a compound 6 column key and it is used as a foreign key in tons of other tables...that was real fun to clean up....use a surrogate 1 column key in that case...but be sure to test....normalize till it hurts then denormalize till it works....I will cover normalization in another post...just wanted to mention it

[1] I know there is text and ntext but those are deprecated

Sizing database files

It has been a while since I wrote some of my best practices posts. I decided to revisit these posts again to see if anything has changed, I also wanted to see if I could add some additional info.

This post will demonstrate that there is a difference in performance if you don't size your database file accordingly. It is a good practice to have your database sized correctly for the next 6 to 12 months, you don't want your server wasting cycles with growing files all the time.

Figure out how big your files are now, figure out how much they will grow in the next year and size your files accordingly, check back every month or so to see if your estimates were correct.

By default SQL Server will create databases with very small files when you create a database and you don't specify the sizes. If you have people creating databases on your servers, consider adding a DDL trigger to notify you when a new DB is added so that you can talk to the database creator and size the files. You also can change the defaults on the server so that you don't have the 10% growth either.

First let's see what the difference is when we have a database where the files will have to grow versus one where the files are big enough for the data that will be inserted.

Here we are creating two databases, one with much bigger files than the other one

The TestBigger database is correctly sized for the data that will be inserted

CREATE DATABASE [TestBigger]
 ON  PRIMARY 
( NAME = N'TestBigger', FILENAME = N'f:\TempTestBigger.mdf' , 
SIZE = 509600KB , FILEGROWTH = 1024KB )
 LOG ON 
( NAME = N'TestBigger_log', FILENAME = N'f:\TempTestBigger_log.ldf' , 
SIZE = 502400KB , FILEGROWTH = 10%)
GO

The TestSmaller database is very small, files will have to be expanded many times to accommodate all the data I will be inserting

CREATE DATABASE [TestSmaller]
 ON  PRIMARY 
( NAME = N'TestSmaller', FILENAME = N'f:\TempTestSmaller.mdf' , 
SIZE = 5280KB , FILEGROWTH = 1024KB )
 LOG ON 
( NAME = N'TestSmaller_log', FILENAME = N'f:\TempTestSmaller_log.ldf' , 
SIZE = 504KB , FILEGROWTH = 10%)
GO

These two stored proc calls are just to verify that the files match with what we specified, you can use sp_helpdb to check the size of a database that you created when you don't specify the file sizes

EXEC sp_helpdb 'TestBigger'

name         filename             filegroup SIZE
TestBigger f:\TempTestBigger.mdf      PRIMARY 509632 KB
TestBigger_log f:\TempTestBigger_log.ldf NULL 502400 KB

EXEC sp_helpdb 'TestSmaller'

name         filename              filegroup SIZE
TestSmaller f:\TempTestSmaller.mdf      PRIMARY 5280 KB
TestSmaller_log f:\TempTestSmaller_log.ldf NULL  512 KB

Next, we are creating two identical tables, one in each database

USE TestSmaller
GO
CREATE TABLE test (SomeName VARCHAR(100), 
SomeID VARCHAR(36), SomeOtherID VARCHAR(100), SomeDate DATETIME)

USE TestBigger
GO
CREATE TABLE test (SomeName VARCHAR(100),
 SomeID VARCHAR(36), SomeOtherID VARCHAR(100), SomeDate DATETIME)

This query is just used so that the data is cached for the two inserts later on, this way the data doesn't have to be fetched from disk for either inserts, you can discard the results after the query is done

USE master
GO


SELECT TOP 1000000 c1.name,NEWID(),NEWID(),GETDATE() 
FROM sys.sysobjects c1
CROSS JOIN sys.sysobjects c2
CROSS JOIN sys.sysobjects c3
CROSS JOIN sys.sysobjects c4

Here is the first insert into the bigger database

INSERT TestBigger.dbo.test
SELECT TOP 1000000 c1.name,NEWID(),NEWID(),GETDATE() 
FROM sys.sysobjects c1
CROSS JOIN sys.sysobjects c2
CROSS JOIN sys.sysobjects c3
CROSS JOIN sys.sysobjects c4

Here is the second insert into the smaller database

INSERT TestSmaller.dbo.test
SELECT TOP 1000000 c1.name,NEWID(),NEWID(),GETDATE() 
FROM sys.sysobjects c1
CROSS JOIN sys.sysobjects c2
CROSS JOIN sys.sysobjects c3
CROSS JOIN sys.sysobjects c4

On several machines I tested on, it takes half the time or less to insert the data in the bigger database compared to the smaller database. On some machines it is almost 5 times faster to insert into the bigger database.

How about on your machine, do you see that the insert into the bigger database takes less than half the time it takes to insert into the smaller database?

Check the sizes of the databases again

EXEC sp_helpdb 'TestBigger'

name         filename            filegroup size
TestBigger f:\TempTestBigger.mdf      PRIMARY 509632 KB
TestBigger_log f:\TempTestBigger_log.ldf NULL 502400 KB

EXEC sp_helpdb 'TestSmaller'

name         filename              filegroup size
TestSmaller f:\TempTestSmaller.mdf      PRIMARY 215296 KB
TestSmaller_log f:\TempTestSmaller_log.ldf NULL 427392 KB

As you can see, the bigger database did not expand, the smaller database expanded a lot.

Autogrow
If you do use autogrow, then make sure you don't use the default 10%, take a look at this message

Date 10/03/2017 12:57:56 PM
Log SQL Server (Current - 11/25/2012 5:00:00 AM)
Source spid62
Message
Autogrow of file 'MyDB_Log' in database 'MyDB' took 104381 milliseconds. Consider using ALTER DATABASE to set a smaller FILEGROWTH for this file.

See that, it took a long time, you don't want to grow a one terabyte file by ten percent, that would be one hundred gigabytes, that is huge. Use something smaller and don't use percent, the bigger the file gets the longer it will take to expand the file.

File placement
Separate the log files from the data files by placing them on separate hard drives. Placing the files on separate drives allows I/O activity to occur at the same time for both the data and log files. Instead of having huge files consider having smaller files in separate filegroups. Put different tables used in the same join queries in different filegroups as well. This will improve performance, because of parallel disk I/O searching for joined data.

Put heavily accessed tables and the nonclustered indexes that belong to those tables on different filegroups. This will improve performance, because of parallel I/O if the files are located on different physical disks. Just remember that you can't separate the clustered indexes from the base table, you can only do this for non clustered indexes. Of course people can get very creative, I have worked with a database once where each table was placed in its own filegroups, there were hundreds of files....what a mess

Tempdb
There are all kinds of recommendations about how many data files you should have for tempdb. Start with 4 files and add more files if you see contention. Paul Randal, has a detailed post here: A SQL Server DBA myth a day: (12/30) tempdb should always have one data file per processor core.

If you can, place tempdb on its own physical drive as well, separated from the user databases.

Consider Solid State hard drives or flash storage for tempdb

See also http://support.microsoft.com/kb/2154845 for recommendations by Microsoft Customer Service and Support

Test, test, test
Never ever blindly follow what you read on the internet, make sure that you test it out first on a QA server before promoting the changes to production!!

Why would anyone use a global temporary table?

SQL Azure has added something called database scoped global temporary tables.

Azure SQL Database supports global temporary tables that are also stored in tempdb and scoped to the database level. This means that global temporary tables are shared for all users’ sessions within the same Azure SQL database. User sessions from other Azure SQL databases cannot access global temporary tables.

They way you add these is by using a double pound sign

They way you create a database scoped global temporary table is like this

CREATE TABLE ##test ( a int, b int);

You are probably thinking...wait a minute, how is this different from a global temporary table? It is pretty much the same but on SQL Azure, it is scoped to the database level, while on prem it is scoped to the instance level


I was listening to the latest SQL Server Radio podcast and the hosts were talking about why anyone would ever need a global temporary table. Why not use a real table instead

I can come up with one answer....

Let's take a look

First create a database and then add a user who has read and write permissions

CREATE DATABASE Test
GO

CREATE LOGIN DenisTest   
    WITH PASSWORD = 'DenisTest', CHECK_EXPIRATION=OFF, CHECK_POLICY=OFF; 

USE Test
GO

CREATE USER DenisTest FOR LOGIN DenisTest;  
GO   

ALTER ROLE [db_datareader] ADD MEMBER DenisTest
GO
ALTER ROLE [db_datawriter] ADD MEMBER DenisTest
GO

Now, let say we want to run a query several times and store the results

USE Test
GO

SELECT name,type_desc,state_desc,size
FROM sys.database_files
WHERE name ='Test'

So I need a table to store the results right?
Let's try creating one......

CREATE TABLE SomeTable (name sysname,
type_desc nvarchar(120),
state_desc nvarchar(120),
size int)

Msg 262, Level 14, State 1, Line 8
CREATE TABLE permission denied in database 'Test'.

That is right, we don't have permissions

Now let's add a double pound sign in front of that table

CREATE TABLE ##SomeTable (name sysname,
type_desc nvarchar(120),
state_desc nvarchar(120),
size int)

And this works

Now we can insert into the table we just created

INSERT ##SomeTable
SELECT name,type_desc,state_desc,size
FROM sys.database_files
WHERE name ='Test'

Also, we can insert and select from other sessions.query windows as well

Of course if the server restarts, this table will be gone. However if you want to capture some query output and look at it, you don't have to wait for someone to give you ddl_admin permission, create the table for you or make you db_owner.


Now why everyone would need a temporary stored procedure...that is another story and I can't really think of a reason