Skip to main content

SQL Server Performance : OR vs UNION ALL

When writing queries, we seem to care less about the performance issue at first. Our first goal is to make the output right. When we get the correct output, we then move on to the next phase we call performance tuning. Today we will try to understand what happens when you write OR in your query. We will do the same thing with UNION ALL and try to understand which one seems to perform better.

                                   Figure: performance boost means winning the race

SQL Server Version

We can follow my instruction on any version of SQL Server.

Database

You don’t need any specific database for this. I shall be creating temp tables and perform queries on those so that you can follow along easily.

Step 1

We create a temp table #User as follows

create table #User (
	Id int,
	[type] int,
	[Name] nvarchar(max),
	CreatedOn Datetime)

This will simply hold Id column, type column which will hold the usertype, Name column, and CreatedOn column. The columns are more or less self-explanatory.

Step 2

After we have created our temp table, I will be adding rows to that table. I will be using  a simple while loop to insert 100,000 rows to this table.

declare @i int;
set @i = 1;

while @i<=100000
begin
	insert into #User values (@i, @i%10, 'User-' + Cast(@i as varchar(max)), GETDATE())
	set @i = @i +1;
end

Here, for Id column, we have inserted incremented value through 1 to 100000. For type column, we have inserted a number from 0 to 9. And for Name column, we have inserted 'User-' followed by the iteration count, So the Name will look like this – 'User-34'

So, our table looks like this –

 Statistics IO

To understand the effect of performance, we will turn on the Statistics IO option like this:

Set Statistics IO ON;

This is a session-based thing, so we need to run it just once.

Include Actual Execution Plan

We will be turning the actual execution plan ON to see the execution plan.

OR [Type]

select Id, [Name] 
from #User 
where [type] = 4 OR [type] = 6
Scan count 1, logical reads 634


UNION ALL [Type]

select Id, [Name] from #User where [type] = 4

UNION ALL

select Id, [Name] from #User where [type] = 6

Scan count 2, logical reads 1268


OR [type and Name]

select Id, [Name] from #User where [type] = 4 OR [Name] ='User-34'

Scan count 1, logical reads 634


UNION ALL [Type and Name]

select Id, [Name] from #User where [type] = 4

UNION ALL

select Id, [Name] from #User where [Name] = 'User-34'

Scan count 2, logical reads 1268


Findings

For the table with no index on it, we get better performance with OR. But in the real-world, we would be using index in our tables most of the time. But for indexed columns, the result can be very different and interesting. Next, we will be looking at how this OR and UNION ALL performs in indexed tables.

Performance Analysis in Skewed Data 

We have seen OR performing better in tables if there are no indexes and the values in the table are more or less non-skewed.

Now, we will be using an indexed version.

Database

We just created an empty Database named “PerformanceTuning”.

USE PerformanceTuning

go

Step 1:

We create a user table with Id (int), Name(varchar) and UserType(int) column  as follows :

CREATE TABLE dbo.[User]( 
    [Id] int,
    [Name] char(200),
    [UserType] int )
GO

Step 2:

We will be inserting some skewed data as follows

INSERT INTO dbo.[User] VALUES (1,'',1)

GO 10000

INSERT INTO dbo.[User] VALUES (2,'',2)

GO 50

INSERT INTO dbo.[User] VALUES (3,'',3)

GO 50

Note that, the 2 and 3 UserType rows are very small compared to UserType 1 data.

Step 3 :

Next, We will be creating clustered index on Id column and Non Clustered index on UserType Column as follows :

CREATE CLUSTERED INDEX CL_Col1 ON dbo.[User] ( Id )

GO

CREATE NONCLUSTERED INDEX IX_Col3 ON dbo.[User] ([UserType])

GO

 

You can use the following script to perform all the stated 3 steps as follows :

USE PerformanceTuning

go

CREATE TABLE dbo.[User]

(             Id int,

              [Name] char(200),

              [UserType] int )

GO

INSERT INTO dbo.[User] VALUES (1,'',1)

GO 10000

INSERT INTO dbo.[User] VALUES (2,'',2)

GO 50

INSERT INTO dbo.[User] VALUES (3,'',3)

GO 50

CREATE CLUSTERED INDEX CL_Col1 ON dbo.[User] ( Id )

GO

CREATE NONCLUSTERED INDEX IX_Col3 ON dbo.[User] ([UserType])

GO

Now, Let's start our performance testing of OR and Union ALL 

OR

We will be using the following query first

Query 1 :

SELECT

     [Name],[UserType]

FROM  dbo.[user]

WHERE UserType = 2

Scan count 1, logical reads 114


We can see that SQL Server has decided to index seek in this particular scenario. Which is great.

Query 2:

SELECT

    [Name],[UserType]

FROM dbo.[user]

WHERE UserType = 2 OR UserType = 3

 

Scan count 1, logical reads 291

We can see that SQL Server has decided to index scan the whole table instead of key lookup, which is kind of frustrating.

 

Union ALL

Now let’s see if Union ALL can be a better alternative to this.

For UNION ALL version, we will be using the following query:

SELECT

     [Name],[UserType]

FROM dbo.[user]

WHERE UserType = 2

UNION ALL

SELECT

    [Name],[UserType]

FROM dbo.[user]

WHERE UserType = 3

If we look at the Message tab for Statistics IO information, we see:

Scan count 2, logical reads 228

Seems better than the OR version, right?

Now, let's look at the execution plan of the query :


 Great, as you can see, SQL Server used index seek for both the part of UNION ALL, meaning SQL Server was less confused this time.

 Conclusion:

  • SQL Server can perform better in OR, if the table has uniform data and there are no indexes in the table
  • SQL Server can perform better in UNION ALL, if the table has skewed data and there are indexes in the table.


Comments

Most Loved Posts

SQL Schema Compare with Visual Studio (A complete Guide)

Introduction When you're working on your Dev Database, an urgent issue comes along, and you instantly solve it by changing Scheme in the Staging Database or Production Database :3, few more these type of patching and you're completely out of sync! A lot of paid alternatives are there like SQL Data Compare by RedGate, but my first choice is Visual Studio's SQL Data Tools. In the following article, I tried to image-describe the steps for SQL Data Tool. Like I said before, there are lots of handly DBAtools out there to compare Schema between two DB Sources. I would like to discuss how you can compare two SQL Server DB with Visual Studio. Make sure you have SQL Server Data tools checked while installing Visual Studio.