Index on a Foreign Key?

About two weeks ago, I had a discussion on Twitter and via email with Jeremiah Peschka (@peschkaj) about placing indexes on Foreign Key columns.  I made the comment that I while indexing a foreign key is not required that I thought they made good candidates.  I also shared this blog post by Greg Low with him. 

Jeremiah mentioned that he had never had SQL Server recommend an index on a foreign key which I thought was strange so I decided to run some tests.  I ran the tests against my copy of AdventureWorks to which I have added data.  Here is the query I ran:

SELECT
    SOD.SalesOrderDetailID,
    SOD.OrderQty,
    SOD.ProductID,
    SOD.UnitPrice,
    SOD.UnitPriceDiscount,
    SOD.LineTotal,
    SOD.UnitPrice2,
    SOH.SalesOrderID,
    SOH.RevisionNumber,
    SOH.OrderDate,
    SOH.DueDate,
    SOH.ShipDate,
    SOH.Status,
    SOH.SalesOrderNumber,
    SOH.PurchaseOrderNumber,
    SOH.AccountNumber,
    SOH.CustomerID,
    SOH.SubTotal,
    SOH.TaxAmt,
    SOH.Freight,
    SOH.TotalDue,
    SOH.Comment
FROM
    Sales.SalesOrderHeader AS SOH JOIN
    Sales.SalesOrderDetail AS SOD ON 
        SOH.SalesOrderID = SOD.SalesOrderID 
WHERE
    SOH.OrderDate >= '12/1/2007'

In my copy of AdventureWorks I have 131465 rows in SalesOrderHeader and 1121317 rows in SalesOrderDetail.  The Max(SalesOrderHeader.OrderDate) is 12/30/2007 so I just queried for the last month (1095 rows).  In this query SalesOrderDetail.SalesOrderID is a foreign key referencing SalesOrderHeader.SalesOrderID and before modification the Clustered Primary Key on SalesOrderDetail is on SalesOrderID, SalesOrderDetailID.  Because SalesOrderID is the lead column in the clustering key the optimizer uses a clustered index seek to access SalesOrderDetail as shown here:

Now, to see if the Optimizer would recommend an index on SalesOrderDetail.SalesOrderID, I altered the clustered primary key on SalesOrderDetail, removing SalesOrderID with the following code:

USE [AdventureWorks]
GO

IF  EXISTS (SELECT * FROM sys.indexes WHERE object_id = OBJECT_ID(N'[Sales].[SalesOrderDetail]') AND name = N'PK_SalesOrderDetail_SalesOrderID_SalesOrderDetailID')
ALTER TABLE [Sales].[SalesOrderDetail] DROP CONSTRAINT [PK_SalesOrderDetail_SalesOrderID_SalesOrderDetailID]
GO

USE [AdventureWorks]
GO

ALTER TABLE [Sales].[SalesOrderDetail] ADD  CONSTRAINT [PK_SalesOrderDetail_SalesOrderID_SalesOrderDetailID] PRIMARY KEY CLUSTERED 
(
    [SalesOrderDetailID] ASC
)WITH (PAD_INDEX  = OFF, STATISTICS_NORECOMPUTE  = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS  = ON, ALLOW_PAGE_LOCKS  = ON) ON [PRIMARY]
GO

I then re-ran my first query and now you can see that I there is a clustered index scan, a hash match join, and, in green at the top, a index recommended on SalesOrderDetail.SalesOrderID:

Next I added the suggested index (remember that you should not just add indexes to production without checking other queries as well):

USE [AdventureWorks]
GO
CREATE NONCLUSTERED INDEX SalesOrderDetails_SalesOrderId
ON [Sales].[SalesOrderDetail] ([SalesOrderID])
INCLUDE ([SalesOrderDetailID],[OrderQty],
        [ProductID],[UnitPrice],
        [UnitPriceDiscount],[LineTotal],
        [UnitPrice2])
GO

Then I re-ran my query, and here’s the execution plan:

Here are the Statistics IO results from the 3 queries as well:

Results with SalesOrderID as lead column in Primary Key (Original State)

Table 'SalesOrderDetail'. Scan count 147, logical reads 516, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Table 'SalesOrderHeader'. Scan count 1, logical reads 461, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Results with SalesOrderID removed from Primary Key and not added as an index.

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 'SalesOrderDetail'. Scan count 1, logical reads 12584, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.


Table 'SalesOrderHeader'. Scan count 1, logical reads 461, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Results with the suggested index on SalesOrderID.

Table 'SalesOrderDetail'. Scan count 147, logical reads 504, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Table 'SalesOrderHeader'. Scan count 1, logical reads 461, physical reads 0, read-ahead reads 0, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

Conclusion

In this case you can see that adding an index on the foreign key, whether the lead column in the Clustered Primary Key or in a non-clustered index, you get much improved performance particularly regarding reads, and that the non-clustered index actual provides a slightly better access path than being part of the clustered primary key.  Certainly you need to test for your individual situations but, I would say that the majority of the time a foreign key will benefit from an index since it is used as a filter in the JOIN criteria.

AdventureWorks can be downloaded from codeplex.  Execution Plans and T-SQL can be downloaded from here.