SQL Server's functions are a valuable addition to TSQL
when used wisely. Jeremiah provides a complete and comprehensive guide
to scalar functions and table-valued functions, and shows how and where
they are best used.
A function, in any programming environment, lets you encapsulate reusable logic and build software that is "composable", i.e.
built of pieces that can be reused and put together in a number of
different ways to meet the needs of the users. Functions hide the steps
and the complexity from other code.
However, in
certain respects, SQL Server's functions are fundamentally different
from functions in other programming environments. In procedural
programming, the piece of functionality that most programmers call a function
should really be called a subroutine, which is more like a miniature
program. These subroutines can go about changing data, introducing side
effects, and generally misbehaving as much as they like.
In SQL Server,
functions adhere much more closely to their mathematic definition of
mapping a set of inputs to a set of outputs. SQL Server's functions
accept parameters, perform some sort of action, and return a result.
They do all of this with no side effects. Nevertheless, in the same way
as subroutines, SQL Server functions can hide complexity from users and
turn a complex piece of code into a re-usable commodity. Functions make
it possible, for example, to create very complex search conditions that
would be difficult and tedious to express in inline T-SQL.
This article describes:
- The
types of user-defined functions (UDFs) that SQL Server supports, both
scalar (which return a single value) and table-valued (which return a
table), and how to use them.
- Some of the more interesting built-in functions
- How and why functions can get you into trouble, and cause terrible performance, if you're not careful about how you use them.
Fun Facts about Functions
This section
describes, briefly, some of the basic characteristics of the various
types of SQL Server function, which we'll explore in more detail as we
progress through the later examples.
As noted in the introduction, all SQL Server functions adhere closely to the mathematic definition of a function i.e. a mapping of inputs to outputs, without have side effects. A function with inputs x and y cannot both return x + y and modify the original value of y. As a matter of fact, that function couldn't even modify y: it is only able to return a new value.
Where Can I Use A Function?
Anywhere!
Well, we can use a
function almost anywhere that we would use a table or column. We can
use a function anywhere that we can use a scalar value or a table.
Functions can be used in constraints, computed columns, joins, WHERE clauses, or even in other functions. Functions are an incredibly powerful part of SQL Server.
Functions can be Scalar or Table-valued
Scalar functions
return a single value. It doesn't matter what type it is, as long as
it's only a single, value rather than a table value. You can use a
scalar function "anywhere that a scalar expression of the same data type is allowed in T-SQL statements" (quote from Books Online). All data types in SQL Server are scalar data types, with the exception of TEXT, NTEXT, ROWVERSION, and IMAGE. Unless you are working with SQL Server 2000, you should be avoiding the TEXT, NTEXT, and IMAGE data types; they are deprecated and will be removed in a future version of SQL Server.
In addition to user-defined scalar functions, SQL Server provides numerous types of built-in scalar functions, some of which we'll cover in more detail later. For example, there are several built-in date functions, such as GETDATE, string functions, such as SUBSTRING, and so on, all of which act on a single value and return a single value. There are also aggregate functions that perform a calculation on a set of values and return a single value, as well as a few ranking functions that produce one row for each input row.
Table-valued
functions (TVFs) return a table instead of a single value. A table
valued function can be used anywhere a table can be used – typically in
the FROM
clause of a query. TVFs make it possible to encapsulate complex logic
in a query. For example, security permissions, calculations, and
business logic can be embedded in a TVF. Careful use of TVFs makes it
easy to create re-usable code frameworks in the database.
One of the
important differences between scalar functions and TVFs is the way in
which they can be handled internally, by the SQL Server query optimizer.
Most developers
will be used to working with compilers that will "inline" trivial
function calls. In other words, in any place where the function is
called, the compiler will automatically incorporate the whole body of
the function into the surrounding code. The alternative is that a
function is treated as interpreted code, and invoking it from the main
body of code requires a jump to a different code block to execute the
function.
The biggest
drawback of SQL Server functions is that they may not be automatically
inlined. For a scalar function that operates on multiple rows, SQL
Server will execute the function once for every row in the result set.
This can have a huge performance impact, as will be demonstrated later
in the article. Fortunately, with TVFs, SQL Server will call them only
once, regardless of the number of rows in the result set and it's often
possible, with a bit of ingenuity, to rewrite scalar functions into
TVFs, and so avoid the row-by-row processing that is inherent with
scalar functions.
In some cases, it
might be necessary to dispense with the TVF altogether, and simply
"manually inline" the function logic into the main code. Of course this
defeats the purpose of creating a function to encapsulate re-usable
logic.
Functions can be Deterministic or Nondeterministic
A deterministic function will return the same
result when it is called with the same set of input parameters. Adding
two numbers together is an example of a deterministic function.
A nondeterministic function, on the other hand,
may return different results every time they are called with the same
set of input values. Even if the state of the data in the database is
the same, the results of the function might be different. The GETDATE
function, for example, is nondeterministic. One caveat of almost all
nondeterministic functions is that they are executed once per statement,
not once per row. If you query 90,000 rows of data and use the RAND
function to attempt to produce a random value for each row you will be
disappointed; SQL Server will only generate a single random number for
the entire statement. The only exception to this rule is NEWID, which will generate a new GUID for every row in the statement.
When we create a function, SQL Server will analyze
the code we've created and evaluate whether the function is
deterministic. If our function makes calls to any nondeterministic
functions, it will, itself, be marked as nondeterministic. SQL Server
relies on the author of a SQL CLR function to declare the function as
deterministic using an attribute.
Deterministic functions can be used in indexed views and computed columns whereas nondeterministic functions cannot.
Keeping things safe: functions can be schema-bound
Functions,
just like views, can be schema bound. Attempts to alter objects that are
referenced by a schema bound function will fail. What does this buy us,
though? Well, just as when schema binding a view, schema binding a
function makes it more difficult to make changes to the underlying data
structures that would break our functions. To create a schema-bound
function we simply specify schema binding as an option during function
creation, as shown in Listing 1.
CREATE FUNCTION Sales.CalculateSalesOrderTotal (@SalesOrderID INT)
RETURNS MONEY
WITH SCHEMABINDING AS
BEGIN
DECLARE @SalesOrderTotal AS MONEY ;
SELECT @SalesOrderTotal =
SUM(sod.LineTotal)
+ soh.TaxAmt
+ soh.Freight
FROM Sales.SalesOrderHeader AS soh
INNER JOIN Sales.SalesOrderDetail AS sod
ON soh.SalesOrderID = sod.SalesOrderID
WHERE soh.SalesOrderID = @SalesOrderId
GROUP BY soh.TaxAmt, soh.Freight ;
RETURN @SalesOrderTotal ;
END;
GO
Listing 1: Creating a Schema Bound Function
Behavior around NULL
We can bind a
function to our database schema, thereby preventing database changes
breaking our function, but what do we do when our function receives NULL
input values? By default, SQL Server will go ahead and run the code in
the function and evaluate all of the parameters passed in, even if one
of those parameters is a NULL value, and so the output of the function is NULL.
This is a waste of processor cycles and we need to avoid this unnecessary work. We could check
every parameter that is passed into a function, but that is a lot of
code to maintain. If you're thinking, "there has to be a better way"
then you're absolutely right. When we create the function we can use the
RETURNS NULL ON NULL INPUT option, which will cause SQL Server to immediately return NULL if any parameters in the function are NULL-valued. Users of SQL Server 2000 and earlier are out of luck, though, as this feature was introduced in SQL Server 2005.
Scalar User-defined Functions
It's time to take a look at some interesting uses
for scalar UDFs, and along the way elucidate the rules that govern how
we create and call them.
Calling Scalar UDFs
There are a few rules that must be followed when creating a function:
- The body of the function must be enclosed in a BEGIN/END block.
- Statements with side effects (insert/update/delete) and
temporary tables may not be used. You can, however, use table variables.
Table variables are allowed in UDFs because they are created as
variables, not through DDL. DDL is viewed as producing a side effect and
is not allowed.
- TRY/CATCH statements are not allowed since CATCH can have the side effect of masking the error state of a given function.
Let's dive straight in and take a look at Listing 2, which shows the code to create, in the AdventureWorks2008 database, a scalar UDF called ProductCostDifference, which will compute the cost difference for a single product, over a time range.
IF OBJECT_ID(N'Production.ProductCostDifference', N'FN') IS NOT NULL
DROP FUNCTION Production.ProductCostDifference ;
GO
CREATE FUNCTION Production.ProductCostDifference
(
@ProductId INT ,
@StartDate DATETIME ,
@EndDate DATETIME
)
RETURNS MONEY
AS
BEGIN
DECLARE @StartingCost AS MONEY ;
DECLARE @CostDifference AS MONEY ;
SELECT TOP 1
@StartingCost = pch.StandardCost
FROM Production.ProductCostHistory AS pch
WHERE pch.ProductID = @ProductId
AND EndDate BETWEEN @StartDate
AND @EndDate
ORDER BY StartDate ASC ;
SELECT TOP 1
@CostDifference = StandardCost - @StartingCost
FROM Production.ProductCostHistory AS pch
WHERE pch.ProductID = @ProductId
AND EndDate BETWEEN @StartDate
AND @EndDate
ORDER BY StartDate DESC ;
RETURN @CostDifference ;
END
Listing 2: Creating a scalar function
Likewise, when we call a UDF, we must follow a few rules:
- quality the function name (e.g. Production.ProductCostDifference) when using a function in a query.
- Optional parameters cannot be omitted, but we can use the DEFAULT keyword to supply the default value.
Listing 3 shows a simple call to our ProductCostDifference function.
SELECT Production.ProductCostDifference(707, '1999-01-01', GETDATE()) ;
/*
column1
----------
1.8504
*/
Listing 3: Executing a user-defined function
Scalar UDFs are a fairly straightforward feature
but there are some drawbacks to them, the biggest one being that, as
discussed earlier, SQL Server has no optimization whereby it can compile
this function as inline code. Therefore, it will simply call it once
for every row to be returned in the result set. Another drawback of
scalar UDFs is that we won't see the true cost of the function when
we're looking at execution plans. This makes it difficult to gauge just
how much a UDF is hurting query performance.
Scalar functions in the SELECT Clause
Listing 4 demonstrates calling our function in the SELECT statement of two simple queries, the only difference being that in the second query we filter out NULL results from our scalar function.
--QUERY 1
SELECT ProductID ,
Name AS ProductName ,
Production.ProductCostDifference
(ProductID, '2000-01-01', GETDATE())
AS CostVariance
FROM Production.Product ;
--QUERY 2
SELECT ProductID ,
Name AS ProductName ,
Production.ProductCostDifference
(ProductID, '2000-01-01', GETDATE())
AS CostVariance
FROM Production.Product
WHERE Production.ProductCostDifference
(ProductID, '2000-01-01', GETDATE())
IS NOT NULL ;
Listing 4: Running a query with a scalar function
Unfortunately, SQL Server is not terribly intelligent in the way that it works with scalar functions. In the first query, our ProductCostDifference function will be executed once for each of the 504 rows in the Production.Product
table. This leads to an increase in disk access, CPU utilization, and
memory utilization. The execution plan for this query is shown in Figure
1.
Figure 1: Execution Plan for Query 1 in Listing 4
We can see that all data is read from disk in the Index Scan operator before being sent to the Compute Scalar
operator, where our function is applied to the data. If we open up the
Properties page for the Compute Scalar node (pressing F4 will do this if
you haven't changed the default SQL Server Management Studio settings)
and examine the Define Values property list.
If this references the function name (rather than the column name), as
it will in this case, then the function is being called once per row.
The situation is even worse for the second query
in Listing 4 in that the function needs to be evaluated twice: once for
every 504 rows in the Production.Product table and once again for the 157 rows that produce a non-NULL result from our scalar function. The execution plan for this query is shown in Figure 2.
Figure 2: Execution Plan for Query 2 in Listing 4
Again, we can establish whether or not a function
is being executed once per row by examining the details of this plan; in
this case, the properties of either the Compute Scalar or the Filter node. The Predicate property of the Filter node shows that that the filter operation is filtering on:
[AdventureWorksCS].[Production].[ProductCostVariance]([AdventureWorksCS].[Production].[Product].[ProductID],'2000-01-01
00:00:00.000',getdate()) IS NOT NULL.
In other words, SQL Server is evaluating the
function once for every row in the product table. No function ‘inlining’
has been performed; we would be able to see the ‘inlined’ source code
if it had been.
This may seem like a trivial point to labor over,
but it can have far reaching performance implications. Imagine that you
have a plot of land. On one side of your plot of land is a box of nails.
How long would it take you to do anything if you only used one nail at a
time and kept returning to the box of nails every time you needed to
use another one? This sort of thing might not be bad for small tasks
like hanging a picture on the wall, but it would become incredibly time
consuming if you were trying to build an addition for your house. The
same thing is happening within your T-SQL. During query evaluation, SQL
Server must evaluate the output of the scalar function once per row.
This could require additional disk access and potentially slow down the
query.
Scalar functions, when used appropriately, can be
incredibly effective. Just be careful to evaluate their use on datasets
similar to the ones you will see in production before you make the
decision to use them; they have some characteristics that may cause
undesirable side effects. If your scalar UDF needs to work on many rows,
one solution is to rewrite it as a table-valued function, as will be
demonstrated a little later.
Scalar functions in the WHERE Clause
Using a scalar function in the WHERE
clause can also have disastrous effects on performance. Although the
symptoms are the same (row-by-row execution), the cause is different.
Consider the call to the built-in scalar function, DATEADD, in Listing 5.
SELECT *
FROM Sales.SalesOrderHeader AS soh
WHERE DATEADD(mm, 12, soh.OrderDate) < GETDATE()
This code will result in a full scan of the Sales.SalesOrderHeader table because SQL Server can't use any index on the OrderDate
column. Instead, SQL Server has to scan every row in the table and
apply the function to each row. A better, more efficient way to write
this particular query would be to move the function, as shown in Listing
6.
SELECT *
FROM Sales.SalesOrderHeader AS soh
WHERE soh.OrderDate < DATEADD(mm, -12, GETDATE())
Listing 6: Better use of a function in the WHERE clause
Optimizing the use of a function in the WHERE
clause isn't always that easy, but in many occasions this problem can
be alleviated through the use of careful design, a computed column, or a
view.
Constraints and Scalar Functions
Functions can be used for more than just
simplifying math; they are also a useful means by which to encapsulate
and enforce rules within the data.
Default Values
Functions can be used to supply the default value
for a column in a table. There is one requirement: no column from the
table can be used as an input parameter to the default constraint.
In Listing 7, we create two tables, Bins and Products, and a user defined function, FirstUnusedProductBin, which will find the first unused bin with the fewest products. We then use the output of the FirstUnusedProductBin function as a default value for the BinID in the Products
table. Creating this default value makes it possible to have a default
storage location for products, which can be overridden by application
code, if necessary.
IF OBJECT_ID(N'dbo.Products', N'U') IS NOT NULL
DROP TABLE dbo.Products ;
GO
IF OBJECT_ID(N'dbo.FirstUnusedProductBin', N'FN') IS NOT NULL
DROP FUNCTION dbo.FirstUnusedProductBin ;
GO
IF OBJECT_ID(N'dbo.Bins', N'U') IS NOT NULL
DROP TABLE dbo.Bins ;
GO
CREATE TABLE dbo.Bins
(
BinID INT IDENTITY(1, 1)
PRIMARY KEY ,
Shelf VARCHAR(2) NOT NULL ,
Bin TINYINT NOT NULL
) ;
CREATE TABLE dbo.Products
(
ProductID INT IDENTITY(1, 1)
PRIMARY KEY ,
ProductName VARCHAR(50) NOT NULL ,
BinID INT REFERENCES dbo.Bins ( BinID )
) ;
GO
CREATE FUNCTION dbo.FirstUnusedProductBin ( )
RETURNS INT
AS
BEGIN
RETURN (SELECT x.BinID
FROM (SELECT b.BinID ,
ROW_NUMBER() OVER
(ORDER BY COUNT(p.ProductID))
AS rn
FROM dbo.Bins AS b
LEFT OUTER JOIN dbo.Products AS p
ON b.BinID = p.BinID
GROUP BY b.BinID
) AS x
WHERE rn = 1
) ;
END
GO
ALTER TABLE dbo.Products ADD CONSTRAINT DF_Products_BinID
DEFAULT dbo.FirstUnusedProductBin() FOR BinID ;
Listing 7: Creating a table with a UDF default
Our UDF will only work for single row INSERTs. We'll explore what happens with a set-based INSERT after we look at the function working correctly with single row INSERTs.
INSERT INTO dbo.Bins (Shelf, Bin)
VALUES ('A', 1),
('A', 2),
('B', 1),
('B', 2),
('B', 3);
INSERT INTO dbo.Products
( ProductName, BinID )
VALUES ( 'widget', DEFAULT ),
( 'sprocket', DEFAULT ),
( 'hammer', DEFAULT ),
( 'flange', DEFAULT ),
( 'gasket', DEFAULT ) ;
GO
-- every bin is full
SELECT *
FROM dbo.Products ;
DELETE FROM dbo.Products
WHERE ProductName = 'hammer' ;
-- bin 3 is empty
SELECT *
FROM dbo.Products ;
INSERT INTO dbo.Products
( ProductName, BinID )
VALUES ( 'plunger', DEFAULT );
-- every bin is full again
SELECT *
FROM dbo.Products ;
Listing 8: Modifying data in the Products table
When we insert data into the Products
table in the first statement it's very easy to see that every bin is
filled. If we remove one product and add a different product, the empty
bin will be re-used.
Although this example demonstrates nicely the way
in which we can use functions to set default values, the implementation
of this function is naïve; once all bins are full it will circle around
and begin adding products to the least full bin. An ideal function would
use a bin-packing algorithm. If you need to use a bin-packing algorithm
in T-SQL, I recommend looking at Chapter 4: Set-based iteration in SQL Server MVP Deep Dives (Kornelis 2009).
What happens if we try to INSERT more than one row at a time?
INSERT INTO dbo.Products
( ProductName
)
SELECT Name
FROM Production.Product ;
SELECT *
FROM dbo.Products ;
Listing 9: Inserting many rows at once
Every row is inserted with the same default value. The FirstUnusedProductBin
function is only called once for the entire transaction. A better way
to enforce a default value that works for both single-row INSERTs and multi-row INSERTs is to use an INSTEAD OF trigger to bypass the set-based INSERT. In effect, we have to force SQL Server to use row-by-row behavior in order to insert a new value in each row.
ALTER TABLE dbo.Bins DROP CONSTRAINT DF_Products_BinID ;
GO
TRUNCATE TABLE dbo.Products ;
GO
DROP TRIGGER TR_Products$Insert ;
GO
CREATE TRIGGER TR_Products$Insert ON dbo.Products
INSTEAD OF INSERT
AS
BEGIN
DECLARE @count INT ;
DECLARE @counter INT ;
SELECT @count = COUNT(*) ,
@counter = 0
FROM inserted ;
SELECT * ,
-- We use a trick with ROW_NUMBER to produce
-- an abitrary, ever increasing row number
-- that is not based on any characteristic of
-- the underlying data.
ROW_NUMBER() OVER ( ORDER BY ( SELECT 1
) ) AS TriggerRowNumber
INTO #inserted
FROM inserted ;
WHILE @counter < @count
BEGIN
INSERT INTO dbo.Products
( ProductName ,
BinID
)
SELECT ProductName ,
dbo.FirstUnusedProductBin()
FROM #inserted
WHERE TriggerRowNumber = @counter + 1 ;
SET @counter = @counter + 1 ;
END
END
INSERT INTO dbo.Products
( ProductName
)
SELECT Name
FROM Production.Product ;
SELECT *
FROM dbo.Products ;
Listing 10: A Multi-row solution
To provide a constantly changing default value for each row we've removed the default constraint and replaced it with an INSTEAD OF trigger for the INSERT.
Unfortunately, this trigger adds significant overhead, but it does
demonstrate the difficulty of using functions to enforce complex default
constraints.
Enforcing Constraints with Functions
Scalar UDFs are often very useful for data
validation and restriction. Many constraints enforce simple, inline
evaluations, such as the "number of federal income tax deductions must
be less than ten"). For example, the CHECK
constraint in Listing 11 enforces the rule that no employee's yearly
bonus is more than 25% of their salary (one could argue that this sort
of salary logic belongs in the application not database, but that debate
is not really relevant to our goal here).
CREATE TABLE dbo.Salaries
(
EmployeeID INT NOT NULL ,
BaseSalary MONEY NOT NULL ,
Bonus MONEY NULL
) ;
ALTER TABLE dbo.Salaries
ADD CONSTRAINT CheckMaxBonus
CHECK ((COALESCE(Bonus, 0) * 4) <= BaseSalary) ;
Listing 11: A simple CHECK constraint
Say, though that we have second rule for this data
says that "no employee may have a salary that is 10 times greater than
the salary of the lowest paid employee". This is a bit trickier because
if we try to add a subquery to the CHECK constraint, we receive an error that "Subqueries are not allowed in this context. Only scalar expressions are allowed."
SQL server is, wisely, preventing us from comparing the output of a
query with a single value. Since we cannot put this comparison inline,
we'll have to create a scalar UDF and use the function to compare the
data.
CREATE FUNCTION dbo.SalaryWithinBounds ( @Salary MONEY )
RETURNS BIT
AS
BEGIN
DECLARE @r_val AS BIT ;
DECLARE @MinSalary AS MONEY ;
SELECT @MinSalary = MIN(BaseSalary)
FROM dbo.Salaries
IF ( @MinSalary * 10 ) > @Salary
SET @r_val = 1
ELSE
SET @r_val = 0
RETURN @r_val ;
END
GO
ALTER TABLE dbo.Salaries
ADD CONSTRAINT CheckMaxSalary
CHECK (dbo.SalaryWithinBounds(BaseSalary) = 1) ;
GO
/* This insert succeeds */
INSERT INTO dbo.Salaries
( EmployeeID, BaseSalary, Bonus )
VALUES ( 5, 1000, 0 ) ;
/* This insert will fail */
INSERT INTO dbo.Salaries
( EmployeeID, BaseSalary, Bonus )
VALUES ( 6, 100000000, 50000 ) ;
Listing 12: Creating a constraint with a function
Functions in constraints are not limited to the
current table; they can reference any table in the database to enforce
data constraints. In the following example we will create two tables – Employees and PayGrades – and implement a CHECK constraint that prevents an employee from having the same or higher pay grade as their manager.
IF OBJECT_ID(N'Employees', N'U') IS NOT NULL
DROP TABLE dbo.Employees ;
GO
IF OBJECT_ID(N'PayGrades', N'U') IS NOT NULL
DROP TABLE dbo.PayGrades ;
GO
IF OBJECT_ID(N'dbo.VerifyPayGrade', N'FN') IS NOT NULL
DROP FUNCTION dbo.VerifyPayGrade ;
GO
CREATE TABLE dbo.PayGrades
(
PayGradeCode CHAR(1) NOT NULL
PRIMARY KEY ,
Position TINYINT NOT NULL
) ;
GO
CREATE TABLE dbo.Employees
(
EmployeeID INT NOT NULL
IDENTITY(1, 1)
PRIMARY KEY ,
ManagerID INT NULL
REFERENCES dbo.Employees ( EmployeeID ) ,
PayGradeCode CHAR(1) NOT NULL
REFERENCES dbo.PayGrades ( PayGradeCode ) ,
FirstName VARCHAR(30) NOT NULL ,
LastName VARCHAR(30) NOT NULL
) ;
GO
CREATE FUNCTION dbo.VerifyPayGrade
(
@ManagerID INT ,
@PayGrade CHAR(1)
)
RETURNS BIT
BEGIN
DECLARE @r AS BIT ,
@ManagerPayGradePosition AS TINYINT ,
@PayGradePosition AS TINYINT ;
SET @r = 0 ;
SELECT @ManagerPayGradePosition = Position
FROM dbo.Employees AS e
INNER JOIN dbo.PayGrades AS pg
ON e.PayGradeCode = pg.PayGradeCode
WHERE e.EmployeeID = @ManagerID ;
SELECT @PayGradePosition = Position
FROM dbo.PayGrades AS pg
WHERE pg.PayGradeCode = @PayGrade ;
SET @r = CASE WHEN @PayGradePosition <
COALESCE(@ManagerPayGradePosition, 999) THEN 1
ELSE 0
END ;
RETURN @r ;
END
GO
ALTER TABLE dbo.Employees
WITH CHECK
ADD
CONSTRAINT CK_Employees_PayGradePosition
CHECK (dbo.VerifyPayGrade(ManagerID, PayGradeCode) = 1) ;
GO
INSERT INTO dbo.PayGrades
( PayGradeCode, Position )
VALUES ( 'A', 10 ),
( 'B', 9 ),
( 'C', 8 ),
( 'D', 7 ),
( 'E', 6 ),
( 'F', 5 ),
( 'G', 4 ),
( 'H', 3 ),
( 'I', 2 ),
( 'J', 1 ) ;
Listing 13: A constraint using functions that access other tables
Having now created our tables and a function to
validate the business rule, we can set about testing that our rules
actually work.
-- These inserts will succeed
INSERT INTO dbo.Employees
( ManagerID, PayGradeCode, FirstName, LastName )
VALUES ( NULL, 'A', 'Kim', 'Abercrombie' ),
( 1, 'B', 'Theodore', 'Stevens' ) ;
GO
-- This insert will fail
INSERT INTO dbo.Employees
( ManagerID ,
PayGradeCode ,
FirstName ,
LastName
)
VALUES ( 2 ,
'B' ,
'James' ,
'Nguyen'
) ;
Listing 14: Verifying our constraints by creating data
The first two inserted rows create the head of the
company at the top pay grade and then we create an immediate
subordinate. The third INSERT
fails because we're attempting to create an employee at the same pay
grade as their manager. When we try to insert a row that violates the
check constraint, an error is returned to the client. We could parse
this error message and return a meaningful message to the client instead
of this:
The
INSERT statement conflicted with the CHECK constraint
"CK_Employees_PayGradePosition". The conflict occurred in database
"AdventureWorks2008", table "dbo.Employees".
Table-valued Functions
Table-valued Functions (TVFs) differ from scalar
functions in that TVFs return an entire table whereas scalar functions
only return a single value. This makes them ideal for encapsulating more
complex logic or functionality for easy re-use. TVFs have the
additional advantage of being executed once to return a large number of
rows (as opposed to scalar functions which must be executed many times
to return many values).
The body of a TVF can either contain just a single
statement or multiple statements, but the two cases are handled very
differently by the optimizer. If the function body contains just a
single statement (often referred to as an "inline TVF"), then the
optimizer treats it in a similar fashion to a view in that it will
"decompose" it and simply reference the underlying objects (there will
be no reference to the function in the resulting execution plan).
However, by contrast, multi-statement TVFs present
an optimization problem for SQL Server; it doesn't know what to do with
them. It treats them rather like a table for which it has no available
statistics – the optimizer assumes that it the TVF will always return
one row. As a result, even a very simple multi-statement TVF can cause
severe performance problems.
Avoiding Row-by-Row Behavior with TVFs
One of the problems with scalar functions is that
they are executed once for every row in the result set. While this is
not a problem for small result sets, it becomes a problem when our
queries return a large number of rows. We can use TVFs to solve this
problem.
We'll start with a relatively simple case of
converting a single-statement scalar function into a single-statement
TVF. We'll then move on to the slightly more complex case of converting
our previous ProductCostDifference
scalar function (Listing 2), which contains multiple statements. As we
discussed earlier, this function can only operate on a single row at a
time. If we wanted to execute that function over several million rows,
we would see a considerable spike in disk I/O and a decrease in
performance.
Single-statement Scalar to Single-statement TVF
Let's take a look at a new example, looking at order data from the AdventureWorks
database. We are specifically interested in the average weight of
orders so we can determine if we need to look into different shipping
options.
Listing 15 creates two functions; the first is a
scalar function and will compute the order weight for any single order.
This is ideal for queries showing the details of a single order or a few
orders, but when it comes to working with a large number of orders this
could cause an incredible amount of disk I/O. The second function is a
single-statement table-valued function that performs the exact same
calculation, but does so over an entire table instead of for a single
row.
IF OBJECT_ID(N'Sales.OrderWeight') IS NOT NULL
DROP FUNCTION Sales.OrderWeight ;
GO
IF OBJECT_ID(N'Sales.tvf_OrderWeight') IS NOT NULL
DROP FUNCTION Sales.tvf_OrderWeight ;
GO
CREATE FUNCTION Sales.OrderWeight ( @SalesOrderID INT )
RETURNS DECIMAL(18, 2)
AS
BEGIN
DECLARE @Weight AS DECIMAL(18, 2) ;
SELECT @Weight = SUM(sod.OrderQty * p.Weight)
FROM Sales.SalesOrderDetail AS sod
INNER JOIN Production.Product AS p
ON sod.ProductID = p.ProductID
WHERE sod.SalesOrderID = @SalesOrderID ;
RETURN @Weight ;
END
GO
CREATE FUNCTION Sales.tvf_OrderWeight ( )
RETURNS TABLE
AS
RETURN
SELECT sod.SalesOrderID ,
SUM(sod.OrderQty * p.Weight) AS OrderWeight
FROM Sales.SalesOrderDetail AS sod
INNER JOIN Production.Product AS p
ON sod.ProductID = p.ProductID
GROUP BY sod.SalesOrderID ;
GO
Listing 15: Two functions for Finding Order Weight
Listing 16 shows the queries for calling each of these functions.
--calling the scalar function
SELECT c.CustomerID ,
AVG(OrderWeight) AS AverageOrderWeight
FROM Sales.Customer AS c
INNER JOIN ( SELECT soh.CustomerID ,
Sales.OrderWeight(soh.SalesOrderID)
AS OrderWeight
FROM Sales.SalesOrderHeader AS soh
WHERE soh.OrderDate BETWEEN '2000-01-01'
AND GETDATE()
) AS x ON c.CustomerID = x.CustomerID
GROUP BY c.CustomerID
ORDER BY c.CustomerID ;
-- calling the single-statement TVF
SELECT c.CustomerID ,
AVG(OrderWeight) AS AverageOrderWeight
FROM Sales.Customer AS c
INNER JOIN Sales.SalesOrderHeader AS soh
ON c.CustomerID = soh.CustomerID
INNER JOIN Sales.tvf_OrderWeight() AS y
ON soh.SalesOrderID = y.SalesOrderID
GROUP BY c.CustomerID
ORDER BY c.CustomerID ;
Listing 16: Using the OrderWeight Functions
Notice the different ways in which the two
functions are invoked. We use our single-statement TVF the same way that
we would use a table. This makes it very easy to use TVFs in our
queries; we only need to join to them and their results will be
incorporated into our existing query.
When you hit the Execute
button for those queries, it should be immediately clear that one of
them, at least, is pretty slow. However, by examining the execution
plans alone, as shown in Figure 6, it's hard to tell which one is the
culprit. In fact, in terms of relative plan cost, you may even conclude
that the first plan is less expensive.
Figure 3: Execution plans for calling the scar function and the TVF
Unfortunately, the top execution plan (for the scalar function), hides any immediately-obvious evidence of the Sales.OrderWeight function reading row-by-row through the Sales.SalesOrderDetail table. Our evidence for that comes, again, from the Compute Scalar operator, where we see direct reference to our Sales.OrderWeight function, indicating that it is being called once per row.
In the second execution plan, for the TVF, we see the index scan against the Sales.SalesOrderDetail
table. Our TVF is called just once to return the required rows, and has
effectively been "inlined"; the plan references only the underlying
objects with no reference to the function itself.
Using SET STATISTICS TIME ON
revealed that (one my machine) the first query executes in 28 seconds
while the second query executes in 0.6 seconds. Clearly the second query
is faster than the first.
Later in the article, we'll take this a step
further and show how to dispense with the TVF altogether and manually
inline the logic of this TVF; a strategy that's sometimes advantageous
from a performance perspective.
Multi-statement Scalar to Multi-statement TVF
Let's now look at the more complex case of converting our multi-statement ProductCostDifference scalar function into a TVF. We're going to start with a straight conversion to a multi-statement TVF.
As noted, a multi-statement TVF is one that
contains more than one statement in the function body. Listing 17 shows a
nearly-direct translation of the original scalar function, in Listing
2, into a multi-statement table-valued function.
IF OBJECT_ID(N'Production.ms_tvf_ProductCostDifference',N'TF' ) IS NOT NULL
--SELECT * FROM sys.objects WHERE name LIKE 'm%'
DROP FUNCTION Production.ms_tvf_ProductCostDifference ;
GO
CREATE FUNCTION Production.ms_tvf_ProductCostDifference
(
@StartDate DATETIME ,
@EndDate DATETIME
)
RETURNS @retCostDifference TABLE
(
ProductId INT ,
CostDifference MONEY
)
AS
BEGIN
DECLARE @workTable TABLE
(
ProductId INT ,
StartingCost MONEY ,
EndingCost MONEY
) ;
INSERT INTO @retCostDifference
( ProductId ,
CostDifference
)
SELECT ProductID ,
StandardCost
FROM ( SELECT pch.ProductID ,
pch.StandardCost ,
ROW_NUMBER() OVER
( PARTITION BY ProductID
ORDER BY StartDate DESC ) AS rn
FROM Production.ProductCostHistory AS pch
WHERE EndDate BETWEEN
@StartDate AND @EndDate
) AS x
WHERE x.rn = 1 ;
UPDATE @retCostDifference
SET CostDifference = CostDifference - StandardCost
FROM @retCostDifference cd
JOIN ( SELECT ProductID ,
StandardCost
FROM ( SELECT pch.ProductID ,
pch.StandardCost ,
ROW_NUMBER() OVER
( PARTITION BY ProductID
ORDER BY StartDate ASC )
AS rn
FROM Production.ProductCostHistory
AS pch
WHERE EndDate BETWEEN
@StartDate AND @EndDate
) AS x
WHERE x.rn = 1
) AS y ON cd.ProductId = y.ProductID ;
RETURN ;
END
Go
Listing 17: A multi-statement TVF
This TVF, Instead of retrieving a single row from
the database and calculating the price difference, pulls back all rows
from the database and calculates the price difference for all rows at
once.
p.Name ,
p.ProductNumber ,
pcd.CostDifference
FROM Production.Product AS p
INNER JOIN Production.ms_tvf_ProductCostDifference
('2001-01-01', GETDATE()) AS pcd
ON p.ProductID = pcd.ProductID ;
Listing 18: Using the TVF
The downside of this multi-statement TVF is that
SQL Server makes the assumption that only one row will be returned from
the TVF, as we can see from the execution plan in Figure 4. With the
data volumes we see in AdventureWorks, this doesn't pose many problems. On a larger production database, though, this would be especially problematic.
Figure 4: Only one row
Multi-statement Scalar to Single-statement TVF
Let's now rewrite our original ProductCostDifference scalar function a second time, this time turning it into a single-statement (or "inline") TVF, as shown in Listing 19.
IF OBJECT_ID(N'Production.tvf_ProductCostDifference') IS NOT NULL
DROP FUNCTION Production.tvf_ProductCostDifference ;
GO
CREATE FUNCTION [Production].[tvf_ProductCostDifference]
(
@StartDate DATETIME ,
@EndDate DATETIME
)
RETURNS TABLE
AS
RETURN
WITH cte
AS ( SELECT pch.ProductID ,
pch.StandardCost AS Cost ,
ROW_NUMBER() OVER ( PARTITION BY pch.ProductID ORDER BY StartDate ASC ) AS rn_1 ,
ROW_NUMBER() OVER ( PARTITION BY pch.ProductID ORDER BY StartDate DESC ) AS rn_2
FROM Production.ProductCostHistory AS pch
WHERE pch.EndDate BETWEEN @StartDate AND @EndDate
)
-- Find the newest price for each product by using
-- grabbing the first (x.rn_1 = 1) row.
-- Then find the last row for each product with
-- x.rn_1 = y.rn_2. Since rn_2 is ordered by StartDate
-- descending, the row in y where rn_2 = 1 is the
-- oldest order.
SELECT x.ProductID ,
y.Cost - x.Cost AS CostDifference
FROM cte AS x
INNER JOIN cte AS y ON x.ProductID = y.ProductID
AND x.rn_1 = y.rn_2
WHERE x.rn_1 = 1 ;
Listing 19: Moving to a single-statement table-valued function
In addition to switching to a table valued
function, we also re-wrote the code to read the table fewer times (two
times, in this case; once for each use of the CTE in Listing 19) by
using two different ROW_NUMBERs. Use of a common table expression also makes it easier and faster to get the oldest and newest row at the same time.
While this initially seems like a lot of
complexity to get to our original goal, it all has a purpose. By
re-writing the TVF to use a common table expression, we can avoid the
performance problem of a multi-statement TVF.
SELECT p.ProductID ,
p.Name ,
p.ProductNumber ,
pcd.CostDifference
FROM Production.Product AS p
INNER JOIN
Production.tvf_ProductCostDifference('2001-01-01',
GETDATE()) AS pcd
ON p.ProductID = pcd.ProductID ;
--QUERY 2
SELECT p.ProductID ,
p.Name ,
p.ProductNumber ,
pcd.CostDifference
FROM Production.Product AS p
INNER JOIN
Production.tvf_ProductCostDifference('2001-01-01',
GETDATE()) AS pcd
ON p.ProductID = pcd.ProductID
WHERE p.Name LIKE 'A%' ;
Listing 20: Two uses of the single-statement TVF
The execution plan for the first query in Listing 20 is shown in Figure 5.
Figure 5: Execution Plan for Query 1 in Listing 20
We can see that the execution plan of the body of Production.tvf_ProductCostDifference
has been "inlined". If the query had not been inlined, we would have
just seen a single operator for executing the table valued function.
Note that there are two scans on ProductCostHistory, because we call the CTE twice in the function, producing two reads of the underlying query).
Again, we can confirm that this TVF is not executed once per row by examining, for example, the Compute Scalar operator, which contains no reference to our function, as well as the the Hash Match
operator. If SQL Server had not inlined the function, we might have
seen a nested loop join to the TVF operator.Looking at the properties of
the Hash Match
node, in Figure 6, we can see that SQL Server not only expects to
perform the join to the body of the TVF just once, but it does perform
that join only once. SQL Server has successfully inlined the TVF and it
is called once for the entire result set, not once for every row.
Figure 6: Properties of the Hash Match node
At this point, we've removed the problem we had
with scalar functions, where they were executed row-by-row. We've also
avoided the problems inherent with multi-statement TVFs. However, we still
may encounter performance problems with these "inline" TVFs. The reason
for this is that the TVF may be evaluated in full before being joined
to our query. This means that we may see several hundred rows returned
from the TVF when our query only returns a few rows. In Listing 20, we
have one query that returns 157 rows and a second query that returns
only a single row. However, turning on STATISTICS IO
reveals that both queries perform the same amount of physical and
logical IO, showing that we're doing the same amount of work whether we
return 157 rows or 1 row. In both cases, we perform two scans of ProductCostHistory at a cost of 10 logical reads.
(157 row(s) affected)
Table 'Product'. Scan
count 0, logical reads 314, physical reads 0, read-ahead reads 0, lob
logical reads 0, lob physical reads 0, lob read-ahead reads 0.
Table 'Worktable'. Scan
count 0, logical reads 0, physical reads 0, read-ahead reads 0, lob
logical reads 0, lob physical reads 0, lob read-ahead reads 0.
Table
'ProductCostHistory'. Scan count 2, logical reads 10, physical reads 0,
read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob
read-ahead reads 0.
(1 row(s) affected)
(1 row(s) affected)
Table 'Product'. Scan
count 0, logical reads 314, physical reads 0, read-ahead reads 0, lob
logical reads 0, lob physical reads 0, lob read-ahead reads 0.
Table 'Worktable'. Scan
count 0, logical reads 0, physical reads 0, read-ahead reads 0, lob
logical reads 0, lob physical reads 0, lob read-ahead reads 0.
Table
'ProductCostHistory'. Scan count 2, logical reads 10, physical reads 0,
read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob
read-ahead reads 0.
(1 row(s) affected)
When you are building your TVFs and the queries
that use them, it's important to remember that the join path may not
always be what you expect. You may come to a point when you have been
optimizing your queries and the performance bottleneck comes down to a
single table-valued function returning a lot of data that is
subsequently filtered down to a few rows. What should you do then?
Rewriting TVFs with CROSS APPLY
In cases where you have a substantial amount of
data that you need to restrict through joins and where conditions, it's
usually better from a performance perspective to dispense with the TVF
altogether and simply "manually inline" the function logic; in other
words, put the body of our T-SQL code inline with the calling code.
Inlining a TVF is as simple as pasting the body of
the function into our query. This can have several benefits. Firstly,
it makes it much easier to determine if our changes are improving
performance. One of the problems of using functions is that
multi-statement TVFs do not report their actual disk I/O to STATISTICS IO,
but they do when you use the SQL Server Profiler. Single statement TVFs
and inlined code will correctly report the disk I/O because it's just
another part of the query.
Inlining TVF code also makes it easier to create
one-off changes for a single query; rather than create a new function,
we can just change the code. When I start inlining code in production I
add a comment in the stored procedure that pointed back to the original
function. This makes it easier to incorporate any improvements that I
may find in the future. Finally, by moving our function to inline code
it's much more likely that SQL Server will make effective join and table
scanning choices and only retrieve the rows that are needed. As
discussed, with TVFs, SQL Server might execute the query in our function
first and return those rows to the outer query before applying any
filtering. If SQL Server returns 200,000 rows and only needs 100, that's
a considerable waste of processing time and disk I/O. However, if SQL
Server can immediately determine the number of rows that will be needed
and which rows will be needed, it will make much more effective querying
choices, including which indexes to use, the type of join to consider,
and the whether or not to perform a parallel operation.
Bear in mind, however, that SQL Server is in
general, very efficient at inlining single-statement TVFs. Listing 21
shows how to manually inline our Sales.tvf_OrderWeight TVF (from Listing 15), by using CROSS APPLY. The CROSS APPLY operator effectively tells SQL Server to invoke a table-valued function for each row returned by an outer query.
AVG(OrderWeight) AS AverageOrderWeight
FROM Sales.Customer AS c
INNER JOIN Sales.SalesOrderHeader AS soh ON c.CustomerID = soh.CustomerID
CROSS APPLY ( SELECT SUM(sod.OrderQty * p.Weight) AS OrderWeight
FROM Sales.SalesOrderDetail AS sod
INNER JOIN Production.Product AS p ON sod.ProductID = p.ProductID
WHERE sod.SalesOrderID = soh.SalesOrderID
GROUP BY sod.SalesOrderID ) AS y
GROUP BY c.CustomerID
ORDER BY c.CustomerID ;
Listing 21: Using CROSS APPLY to inline the OrderWeight TVF
However, we'll see no difference in the execution
plan compared to the one we saw from calling the TVF (Figure 3). One of
the benefits of using single statement TVFs is that SQL Server is
frequently able to optimize the queries and inline the TVF for us. By
knowing how to write optimal TVFs we can build reusable code and take
advantage of SQL Server's ability to automatically inline
well-constructed TVFs.
We can also manually inline the logic of our multi-statemnt TVFs. Listing 22 shows how to do this for our tvf_ProductCostDifference
function. However, again, we don't get additional benefit in this
particular case. It turns out that attempting to inline the common table
expression has similar I/O implications to calling the function.
Regardless of whether we leave the TVF as-is or inline the function
logic using just the CTE, SQL Server will have to enter the CTE and
evaluate the CTE query twice, once for each join to the CTE, and we'll
see the same amount of physical I/O in each case.
SET STATISTICS IO ON ;
DECLARE @StartDate AS DATETIME ;
DECLARE @EndDate AS DATETIME ;
SET @StartDate = '2001-01-01' ;
SET @EndDate = GETDATE() ;
-- calling the TVF
SELECT p.ProductID ,
p.Name ,
p.ProductNumber ,
pcd.CostDifference
FROM Production.Product AS p
INNER JOIN Production.tvf_ProductCostDifference
(@StartDate, @EndDate)
AS pcd ON p.ProductID = pcd.ProductID
WHERE p.Name LIKE 'A%' ;
-- manually inlining the TVF logic
WITH cte
AS ( SELECT pch.ProductID ,
pch.StandardCost AS Cost ,
ROW_NUMBER() OVER
( PARTITION BY pch.ProductID
ORDER BY StartDate ASC ) AS rn_1 ,
ROW_NUMBER() OVER
( PARTITION BY pch.ProductID
ORDER BY StartDate DESC ) AS rn_2
FROM Production.ProductCostHistory AS pch
WHERE pch.EndDate BETWEEN @StartDate AND @EndDate
)
SELECT x.ProductID ,
y.Cost - x.Cost AS CostDifference
FROM Production.Product AS p
INNER JOIN cte AS x ON p.ProductID = x.ProductID
INNER JOIN cte AS y ON x.ProductID = y.ProductID
AND x.rn_1 = y.rn_2
WHERE x.rn_1 = 1
AND p.Name LIKE 'A%'
Listing 22: Re-writing a TVF with an inline function body'
If you run Listing 22, you'll find that the
execution plan for the manually-inlined code is very similar (but not
identical) to the plan for calling the TVF (Figure 5) and performs
similar I/O.
Best Practices for TVFs
Table-valued functions are best used when you will
be performing operations on a large number of rows at once. Typically
this will be something that could be accomplished through a complex
subquery or functionality that you will re-use multiple times in your
database. TVFs should be used when you can always work with the same set
of parameters – dynamic SQL is not allowed within functions in SQL
Server.
It's best to use TVFs when you only have a small
dataset that could be used in the TVF. Once you start getting into
larger numbers of rows, TVFs can become very slow since all the results
of the TVF query are evaluated before being filtered by the outer query.
When this starts to happen it is best to inline the code. If inlining
the TVF code doesn't work, you can even look into re-writing the query
slightly to use a JOIN instead of a CROSS APPLY. This might complicate the query, but it can lead to dramatic performance improvements.
A few useful Built-in Functions
To finish off this article, we'll briefly review some of the built-in scalar functions that I've frequently found useful.
COALESCE
COALESCE takes an unlimited list of arguments and returns the first non-NULL expression. One of the advantages of COALESCE is that it can be used to replace lengthy CASE statements.
SELECT CASE WHEN col_1 IS NOT NULL THEN col_1
WHEN col_2 IS NOT NULL THEN col_2
WHEN col_3 IS NOT NULL THEN col_3
END AS result
Listing 23: A CASE statement
The code in Listing 23 can be simplified by using COALESCE.
SELECT COALESCE(col_1, col_2, col_3) AS result
Listing 24: COALESCE
DATEADD and DATEDIFF
The best way to modify date and time values is by using the DATEADD and DATEDIFF functions. The DATEADD function can be used to add or subtract an interval to part of a date.
SELECT DATEADD(hh, 5, GETDATE()) ;
SELECT DATEADD(hh, -5, GETDATE()) ;
SELECT DATEADD(dd, 1, GETDATE()) ;
Listing 25: Modifying dates
The DATEDIFF function can be used to calculate the difference between to dates. DATEDIFF is similar to DATEADD - DATEDIFF gets the difference between two dates using the given time interval (year, months, seconds, and so on).
SELECT DATEDIFF(dd, '2010-02-05', GETDATE()) ;
SELECT DATEDIFF(hh, '2010-01-04 09:37:00', '2010-03-02 17:54:25') ;
Listing 26: Finding the difference between two points in time
One practical use of the DATEDIFF function is to find the beginning of the current day or month.
SELECT DATEADD(dd, DATEDIFF(dd, 0, GETDATE()), 0) AS beginning_of_day ;
SELECT DATEADD(mm, DATEDIFF(mm, 0, GETDATE()), 0)
AS beginning_of_month ;
Listing 27: Beginning of the day or month
This approach to getting the beginning of the day
computes the number of days since the dawn of SQL Server time (January
1, 1900). We then add that number of days to the dawn of SQL Server time
and we now have to beginning of the current hour, day, month, or even
year.
SIGN
SIGN returns +1, 0, or -1 based on the expression supplied.
Listing 28: Using SIGN
STUFF
STUFF
is a powerful built-in function. It inserts one string into another. In
addition, it also removes a specific number of characters from one
string and adds the second string in place of the removed characters.
That isn't a very clear explanation, so let's take a look at an example.
SELECT STUFF('junk goes here', 1, 4, 'STUFF')
/*
---------------
STUFF goes here
*/
Listing 29: Using STUFF
The STUFF function can be combined with several XML functions to create a comma-separated list of values from a table.
SELECT STUFF(( SELECT ',' + Name AS [text()]
FROM Production.Culture
FOR
XML PATH('')
), 1, 1, '') AS cultures ;
Listing 30: Creating a comma-separated list with STUFF
Summary
User-defined functions give us the ability to
create reusable chunks of code that simplify the code we write. UDFs can
be embedded in queries as single, scalar values or as table valued
functions. Effective use of UDFs can increase the readability of your
code, enhance functionality, and increase maintainability.
