Sometimes data in a data warehouse can get duplicated. There are a lot of reasons for this, including but not limited to:
The data pull may have a join that makes duplicating data unavoidable. There may be an error in the source system. There might be duplicate data in a staging table, and nobody wrote a check for that. There might be no master data management in place.
Whatever the reason, having duplicate data in your data warehouse means your numbers may not add up right. The best place to get rid of dups is in the ETL process, but sometimes you find mistakes in data after you load it to prod. When you are confident in your code, and you are sure any duplicates are in fact duplicate records from a business perspective, you will need a way to get rid of them. The technique below utilizes the ROW_NUMBER() function.
This example was developed and tested on SQL Server 2016 Development Edition.
Examples
Extensive Duplicate Record Removal Process
Let’s create our tables. We’re going to create two tables: one fact table and one dimension table.
Run the following script.
In [ ]:
USE demo
IF EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.FactProductSales') AND type in (N'U'))
ALTER TABLE dbo.FactProductSales DROP CONSTRAINT IF EXISTS FK_ProductSales_Products
GO
DROP TABLE IF EXISTS dbo.FactProductSales
GO
--Create our product table
IF EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.DimProducts') AND type in (N'U'))
ALTER TABLE dbo.DimProducts DROP CONSTRAINT IF EXISTS FK_Products_Products
GO
DROP TABLE IF EXISTS dbo.DimProducts
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
IF NOT EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.DimProducts') AND type in (N'U'))
BEGIN
CREATE TABLE dbo.DimProducts(
ProductID bigint IDENTITY(1,1) NOT NULL,
ProductName nvarchar(50) NULL,
CONSTRAINT PK_Products PRIMARY KEY CLUSTERED
(
ProductID ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]
END
GO
IF NOT EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_Products_Products') AND parent_object_id = OBJECT_ID(N'dbo.DimProducts'))
ALTER TABLE dbo.DimProducts WITH CHECK ADD CONSTRAINT FK_Products_Products FOREIGN KEY(ProductID)
REFERENCES dbo.DimProducts (ProductID)
GO
IF EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_Products_Products') AND parent_object_id = OBJECT_ID(N'dbo.DimProducts'))
ALTER TABLE dbo.DimProducts CHECK CONSTRAINT FK_Products_Products
GO
--Create our product sales table
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
IF NOT EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.FactProductSales') AND type in (N'U'))
BEGIN
CREATE TABLE dbo.FactProductSales(
InvoiceID bigint IDENTITY(1,1) NOT NULL,
ProductID bigint NOT NULL,
Price money NOT NULL,
ItemCount int NOT NULL,
InvoiceTotal AS (Price*ItemCount),
CONSTRAINT PK_ProductSales PRIMARY KEY CLUSTERED
(
InvoiceID ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]
END
GO
IF NOT EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_ProductSales_Products') AND parent_object_id = OBJECT_ID(N'dbo.FactProductSales'))
ALTER TABLE dbo.FactProductSales WITH CHECK ADD CONSTRAINT FK_ProductSales_Products FOREIGN KEY(ProductID)
REFERENCES dbo.DimProducts (ProductID)
GO
IF EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_ProductSales_Products') AND parent_object_id = OBJECT_ID(N'dbo.FactProductSales'))
ALTER TABLE dbo.FactProductSales CHECK CONSTRAINT FK_ProductSales_Products
GO
Expand the database in the object explorer and take a look at the tables that got created. One table called DimProducts holds our products and the other table called FactProductSales holds the sales transactions of those items.
Now let’s create some sample data. Run this script.
In [ ]:
USE demo
INSERT INTO DimProducts(ProductName)
SELECT 'Microsoft Office 365'
UNION ALL
SELECT 'Microsoft Access 2013'
UNION ALL
SELECT 'Microsoft Access 2013'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft Visual Studio 2013'
UNION ALL
SELECT 'Microsoft Visual Studio 2013'
--let's check our work
SELECT * FROM DimProducts
--Now let's load some sales
DECLARE @TempSales TABLE (Price MONEY, ItemCount INT)
INSERT INTO @TempSales
SELECT 300.00, 5
UNION
SELECT 520.00, 10
--Each record in @TempSales doesn't represent a transaction.
--We just want to have more than one invoice for a particular
--product.
INSERT INTO FactProductSales(ProductID, Price, ItemCount)
SELECT p.ProductID, ts.Price, ts.ItemCount
FROM DimProducts p
CROSS JOIN @TempSales ts
--let's check our work
SELECT * FROM FactProductSales ORDER BY ProductID
--And with the join
SELECT p.ProductName, ps.ProductID, ps.InvoiceID AS InvoiceNumber, ps.Price, ps.ItemCount, ps.InvoiceTotal
FROM DimProducts p
JOIN FactProductSales ps
ON p.ProductID = ps.ProductID
ORDER BY p.ProductName
This generates three outputs so you can see the results of the table load operation. The first output is from the products dimension table where you can see that there are duplicates in the data. The second is from the product sales fact table. If you’re sharp, you can see that there are fact records attached to the duplicate dimension records. The final query shows the results of a join between the two tables where you can again see the impact of having duplicate dimension records.
Since we know that there are duplicate records in the dimension table, we now have to get rid of those records, but we cannot do it by simply deleting them. First, we have to reassign those fact records to the prime dimension record, then we can blow away the unnecessary dimension records. Picking which dimension record is the prime dimension record is complex and beyond the scope of this tutorial. For now, we’ll just choose an arbitrary record and call that the prime record.
There are two key SQL objects that we’re going to use to make this change happen: CTEs, and the ROW_NUMBER() function. These are both complex objects and merit further discussion. For now, just go on faith and follow this example. When you get a minute, you can read the TSQL documentation. CTEs are fairly common, but the ROW_NUMBER() function might be new to you.
The process of de-duplication works like this. First, we group and number the records in the dimension table. The numbering will tell you if there is more than one instance of a specific record. We use ROW_NUMBER() to assign ordinal values to the dimension records and group the counting by ProductName. We load these records into a table variable, so they are easy to work with. Any records with a Row Number value of 1 are original records. Any record with a Row Number value greater than 1 is a duplicate.
The next step is to pair duplicate record IDs with prime record IDs. We do this so we can reassign the duplicate record IDs in the fact table to the proper prime dimension record. We do that by being clever with some table variables, a CTE, and a join.
Now we can finally fix our tables. From here, it’s a pretty straight-forward process to update our fact table with the proper IDs. Once that is done, we’re clear to blow away our duplicates in our dimension table. Run the script below.
In [ ]:
USE demo
--Looks peachy but we know some of those product IDs are bogus.
--Let's fix it.
--First we need keep track of the dups
--If you have a large amount of dups, use a
--temp table instead.
DECLARE @DuplicateProductIDs TABLE(RowNumber INT, ProductID BIGINT, ProductName NVARCHAR(50))
INSERT INTO @DuplicateProductIDs(RowNumber, ProductID, ProductName)
SELECT ROW_NUMBER() OVER (PARTITION BY ProductName ORDER BY ProductName) AS RowNumber, ProductID, ProductName
FROM DimProducts
SELECT * FROM @DuplicateProductIDs
--Let's fix our child table so the records point to the right product
--First we need to transform the Dup ID info so we know which dup ID
--belongs to the original ID
DECLARE @RealProductIDsWithDupes TABLE(OriginalProductID BIGINT, DuplicateProductID BIGINT, ProductName NVARCHAR(50))
;
WITH OriginalProductIDs(OriginalProductID, ProductName)
AS(
SELECT ProductID, ProductName
FROM @DuplicateProductIDs
WHERE RowNumber = 1
),
DuplicateProductIDs(DuplicateProductID, ProductName)
AS(
SELECT ProductID, ProductName
FROM @DuplicateProductIDs
WHERE RowNumber <> 1
)
INSERT INTO @RealProductIDsWithDupes(OriginalProductID, DuplicateProductID, ProductName)
SELECT opid.OriginalProductID, dpid.DuplicateProductID, opid.ProductName
FROM OriginalProductIDs opid
JOIN DuplicateProductIDs dpid
ON opid.ProductName = dpid.ProductName
--check
SELECT * FROM @RealProductIDsWithDupes
--Now let's fix our child table
--Before
SELECT * FROM FactProductSales ORDER BY ProductID
UPDATE ps
SET ps.ProductID = rpids.OriginalProductID
FROM FactProductSales ps
JOIN @RealProductIDsWithDupes rpids
ON rpids.DuplicateProductID = ps.ProductID
--After
--Number of rows should not have changed but the
--number of distinct product IDs should be reduced.
SELECT * FROM FactProductSales ORDER BY ProductID
--Now that we won't violate any FK restraints,
--we can safely blow away dups in the parent
DELETE
FROM DimProducts
WHERE ProductID IN(
SELECT ProductID
FROM @DuplicateProductIDs
WHERE RowNumber <> 1
)
--Check
SELECT * FROM DimProducts
--Final result
SELECT ps.InvoiceID AS InvoiceNumber, ps.ProductID, p.ProductName, ps.Price, ps.ItemCount, ps.InvoiceTotal
FROM DimProducts p
JOIN FactProductSales ps
ON p.ProductID = ps.ProductID
ORDER BY p.ProductID
--Clean up your mess
--DROP TABLE FactProductSales
--DROP TABLE DimProducts
This was a complex example. Ideally, you want to get rid of duplicate records early in the pipeline. The most common case I find is getting rid of them in the staging tables. However, this is not always possible.
The techniques taught here can be generalized to other use cases. Below is the example code in its entirety.
In [ ]:
USE demo
--Create our product table
IF EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.DimProducts') AND type in (N'U'))
ALTER TABLE dbo.DimProducts DROP CONSTRAINT IF EXISTS FK_Products_Products
GO
DROP TABLE IF EXISTS dbo.DimProducts
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
IF NOT EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.DimProducts') AND type in (N'U'))
BEGIN
CREATE TABLE dbo.DimProducts(
ProductID bigint IDENTITY(1,1) NOT NULL,
ProductName nvarchar(50) NULL,
CONSTRAINT PK_Products PRIMARY KEY CLUSTERED
(
ProductID ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]
END
GO
IF NOT EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_Products_Products') AND parent_object_id = OBJECT_ID(N'dbo.DimProducts'))
ALTER TABLE dbo.DimProducts WITH CHECK ADD CONSTRAINT FK_Products_Products FOREIGN KEY(ProductID)
REFERENCES dbo.DimProducts (ProductID)
GO
IF EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_Products_Products') AND parent_object_id = OBJECT_ID(N'dbo.DimProducts'))
ALTER TABLE dbo.DimProducts CHECK CONSTRAINT FK_Products_Products
GO
--Create our product sales table
GO
IF EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.FactProductSales') AND type in (N'U'))
ALTER TABLE dbo.FactProductSales DROP CONSTRAINT IF EXISTS FK_ProductSales_Products
GO
DROP TABLE IF EXISTS dbo.FactProductSales
GO
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
IF NOT EXISTS (SELECT * FROM sys.objects WHERE object_id = OBJECT_ID(N'dbo.FactProductSales') AND type in (N'U'))
BEGIN
CREATE TABLE dbo.FactProductSales(
InvoiceID bigint IDENTITY(1,1) NOT NULL,
ProductID bigint NOT NULL,
Price money NOT NULL,
ItemCount int NOT NULL,
InvoiceTotal AS (Price*ItemCount),
CONSTRAINT PK_ProductSales PRIMARY KEY CLUSTERED
(
InvoiceID ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]
END
GO
IF NOT EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_ProductSales_Products') AND parent_object_id = OBJECT_ID(N'dbo.FactProductSales'))
ALTER TABLE dbo.FactProductSales WITH CHECK ADD CONSTRAINT FK_ProductSales_Products FOREIGN KEY(ProductID)
REFERENCES dbo.DimProducts (ProductID)
GO
IF EXISTS (SELECT * FROM sys.foreign_keys WHERE object_id = OBJECT_ID(N'dbo.FK_ProductSales_Products') AND parent_object_id = OBJECT_ID(N'dbo.FactProductSales'))
ALTER TABLE dbo.FactProductSales CHECK CONSTRAINT FK_ProductSales_Products
GO
INSERT INTO DimProducts(ProductName)
SELECT 'Microsoft Office 365'
UNION ALL
SELECT 'Microsoft Access 2013'
UNION ALL
SELECT 'Microsoft Access 2013'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft SQL Server 2012'
UNION ALL
SELECT 'Microsoft Visual Studio 2013'
UNION ALL
SELECT 'Microsoft Visual Studio 2013'
--let's check our work
SELECT * FROM DimProducts
--Now let's load some sales
DECLARE @TempSales TABLE (Price MONEY, ItemCount INT)
INSERT INTO @TempSales
SELECT 300.00, 5
UNION
SELECT 520.00, 10
--Each record in @TempSales doesn't represent a transaction.
--We just want to have more than one invoice for a particular
--product.
INSERT INTO FactProductSales(ProductID, Price, ItemCount)
SELECT p.ProductID, ts.Price, ts.ItemCount
FROM DimProducts p
CROSS JOIN @TempSales ts
--let's check our work
SELECT * FROM FactProductSales ORDER BY ProductID
--And with the join
SELECT p.ProductName, ps.ProductID, ps.InvoiceID AS InvoiceNumber, ps.Price, ps.ItemCount, ps.InvoiceTotal
FROM DimProducts p
JOIN FactProductSales ps
ON p.ProductID = ps.ProductID
ORDER BY p.ProductName
--Looks peachy but we know some of those product IDs are bogus.
--Let's fix it.
--First we need keep track of the dups
--If you have a large amount of dups, use a
--temp table instead.
DECLARE @DuplicateProductIDs TABLE(RowNumber INT, ProductID BIGINT, ProductName NVARCHAR(50))
INSERT INTO @DuplicateProductIDs(RowNumber, ProductID, ProductName)
SELECT ROW_NUMBER() OVER (PARTITION BY ProductName ORDER BY ProductName) AS RowNumber, ProductID, ProductName
FROM DimProducts
SELECT * FROM @DuplicateProductIDs
--Let's fix our child table so the records point to the right product
--First we need to transform the Dup ID info so we know which dup ID
--belongs to the original ID
DECLARE @RealProductIDsWithDupes TABLE(OriginalProductID BIGINT, DuplicateProductID BIGINT, ProductName NVARCHAR(50))
;
WITH OriginalProductIDs(OriginalProductID, ProductName)
AS(
SELECT ProductID, ProductName
FROM @DuplicateProductIDs
WHERE RowNumber = 1
),
DuplicateProductIDs(DuplicateProductID, ProductName)
AS(
SELECT ProductID, ProductName
FROM @DuplicateProductIDs
WHERE RowNumber <> 1
)
INSERT INTO @RealProductIDsWithDupes(OriginalProductID, DuplicateProductID, ProductName)
SELECT opid.OriginalProductID, dpid.DuplicateProductID, opid.ProductName
FROM OriginalProductIDs opid
JOIN DuplicateProductIDs dpid
ON opid.ProductName = dpid.ProductName
--check
SELECT * FROM @RealProductIDsWithDupes
--Now let's fix our child table
--Before
SELECT * FROM FactProductSales ORDER BY ProductID
UPDATE ps
SET ps.ProductID = rpids.OriginalProductID
FROM FactProductSales ps
JOIN @RealProductIDsWithDupes rpids
ON rpids.DuplicateProductID = ps.ProductID
--After
--Number of rows should not have changed but the
--number of distinct product IDs should be reduced.
SELECT * FROM FactProductSales ORDER BY ProductID
--Now that we won't violate any FK restraints,
--we can safely blow away dups in the parent
DELETE
FROM DimProducts
WHERE ProductID IN(
SELECT ProductID
FROM @DuplicateProductIDs
WHERE RowNumber <> 1
)
--Check
SELECT * FROM DimProducts
--Final result
SELECT ps.InvoiceID AS InvoiceNumber, ps.ProductID, p.ProductName, ps.Price, ps.ItemCount, ps.InvoiceTotal
FROM DimProducts p
JOIN FactProductSales ps
ON p.ProductID = ps.ProductID
ORDER BY p.ProductID
--Clean up your mess
DROP TABLE FactProductSales
DROP TABLE DimProducts
