Optimize Large SQL Server Insert, Update and Delete Processes by Using Batches

By: Eduardo Pivaral |  |   Related Tips:More >T-SQL

Problem

Sometimes you must perform DML processes (insert, update, delete or combinations of these) on large SQL Server tables. If your database has a high concurrency these types of processes can lead toblocking or filling up the transaction log , even if you run these processes outside of business hours. So maybe you were tasked to optimize some processes to avoid large log growths and minimize locks on tables. How can this be done?

Solution

We will do these DML processes using batches with the help of @@ROWCOUNT .  This also give you the ability to implement custom “stop-resume” logic. We will show you a general method, so you can use it as a base to implement your own processes.

Please note that we will not focus onindexes on this tip, of course this can help queries, but I want to show you a worst-case scenario and index creation is another topic.

Basic algorithm

The basic batch process is something like this:

DECLARE @id_control INT
DECLARE @batchSize INT
DECLARE @results INT

SET @results = 1 --stores the row count after each successful batch
SET @batchSize = 10000 --How many rows you want to operate on each batch
SET @id_control = 0 --current batch 

-- when 0 rows returned, exit the loop
WHILE (@results > 0) 
BEGIN
   -- put your custom code here
   SELECT * -- OR DELETE OR UPDATE
   FROM <any Table>
   WHERE <your logic evaluations>
   (
      AND <your PK> > @id_control
      AND <your PK> <= @id_control + @batchSize
   )
   -- very important to obtain the latest rowcount to avoid infinite loops
   SET @results = @@ROWCOUNT

   -- next batch
   SET @id_control = @id_control + @batchSize
END

To explain the code, we use a WHILE loop and run our statements inside the loop and we set a batch size (numeric value) to indicate how many rows we want to operate on each batch.

For this approach, I am assuming the primary key is either an int or a numeric data type, so for this algorithm to work you will need that type of key. So for alphanumeric or GUID keys, this approach won't work, but you can implement some other type of custom batch processing with some additional coding.

So, with the batch size and the key control variable, we validate the rows in the table are within the range.

Important Note:Your process will need to always operate on at least some rows in each batch.  If a batch does not operate on any rows, the process will end as row count will be 0. If you have a situation where only some rows from a large table will be affected, it is better and more secure to use the index/single DML approach. Another approach for these cases is to use a temporary table to filter the rows to be processed and then use this temp table in the loop to control the process.

Our example setup

We will use a test table [MyTestTable] with this definition:

CREATE TABLE [dbo].[MyTestTable](
   [id] [bigint] IDENTITY(1,1) NOT NULL,
   [dataVarchar] [nvarchar](50) NULL,
   [dataNumeric] [numeric](18, 3) NULL,
   [dataInt] [int] NULL,
   [dataDate] [smalldatetime] NULL,
 CONSTRAINT [PK_MyTestTable] PRIMARY KEY CLUSTERED 
(
   [id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]
GO

It contains random information and 6,000,000 records.

Executing a SELECT statement

Here we execute a simple SELECT statement over the entire table. Note, I enabled statistics IO and cleared the data cache first so we have better results for comparison.

DBCC DROPCLEANBUFFERS 

SET STATISTICS IO ON

SELECT *
FROM [dbo].[MyTestTable]
WHERE dataInt > 600

These are the IO results:

Table 'MyTestTable'. Scan count 1, logical reads 65415, physical reads 2, read-ahead reads 65398, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

The SELECT took 1:08 minutes and retrieved 2,395,317 rows.

SELECT Statement using batches

For the same SELECT we implement the following process to do it in batches:

DBCC DROPCLEANBUFFERS 

SET STATISTICS IO ON

DECLARE @id_control INT
DECLARE @batchSize INT
DECLARE @results INT

SET @results = 1
SET @batchSize = 100000
SET @id_control = 0

WHILE (@results > 0)
BEGIN
   -- put your custom code here
   SELECT *
   FROM [dbo].[MyTestTable]
   WHERE dataInt > 600
   AND id > @id_control
      AND id <= @id_control + @batchSize
   
   -- very important to obtain the latest rowcount to avoid infinite loops
   SET @results = @@ROWCOUNT

   -- next batch
   SET @id_control = @id_control + @batchSize
END

The IO results (for each batch):

Table 'MyTestTable'. Scan count 1, logical reads 1092, physical reads 0, read-ahead reads 1088, lob logical reads 0, lob physical reads 0, lob read-ahead reads 0.

If we multiply it for 60 batches performed it should be around 65,500 logical reads (approximately the same as before, this makes sense since is the same data we are accessing).

But if we look at the overall execution time, it improves by around 10 seconds, with the same number of rows:

A SELECT statement is probably not the best way to demonstrate this, so let's proceed with an UPDATE statement.

UPDATE Statement using batches

We will do an UPDATE on a varchar field with random data (so our test is more real), after clearing the cache, we will execute the code.

This is a screenshot of the transaction log before the operation.

DBCC DROPCLEANBUFFERS 

BEGIN TRAN;

UPDATE [dbo].[MyTestTable]
SET dataVarchar = N'Test UPDATE 1'
WHERE dataInt > 200;

COMMIT TRAN;

The execution took 37 seconds on my machine.

To find the rows affected, we perform a simple count and we get 4,793,808 rows:

SELECT COUNT(1)
FROM [dbo].[MyTestTable]
WHERE dataVarchar = N'Test UPDATE 1'

Checking the log size again, we can see it grew to 1.5 GB (and then released the space since the database is in SIMPLE mode):

Let's proceed to execute the same UPDATE statement in batches.  We will just change the text Test UPDATE 1 for Test UPDATE 2 , this time using the batch process. I also shrunk the transaction log to its original size and perform a cache cleanup before executing.

DBCC DROPCLEANBUFFERS
 
DECLARE @id_control INT
DECLARE @batchSize INT
DECLARE @results INT

SET @results = 1
SET @batchSize = 1000000
SET @id_control = 0

WHILE (@results > 0)
BEGIN
   -- put your custom code here
   BEGIN TRAN;

   UPDATE [dbo].[MyTestTable]
   SET dataVarchar = N'Test UPDATE 2'
   WHERE dataInt > 200
   AND id > @id_control
   AND id <= @id_control + @batchSize

   -- very important to obtain the latest rowcount to avoid infinite loops
   SET @results = @@ROWCOUNT

   COMMIT TRAN;
   
   -- next batch
   SET @id_control = @id_control + @batchSize
END

This time the query execution was 18 seconds, so there was an improvement in time.

As we can see there was an improvement with the log spaced used.  This time the log grew to 0.43 GB.

The last thing to verify is the number of rows affected. We can see we have the same row count as the UPDATE above, 4,793,808 rows.

As you can see, for very large DML processes, running in smaller batches can help on execution time and transaction log use.

The only drawback of this method is that your key must be a sequential number and there must ne at least one row in each batch, so the process does not end before being applied to all data.

Next Steps

• You can determine if your process can use this batch method just running the SELECT statements and comparing the number of expected rows with the results.
• You can increase/decrease the batch size to suit your needs, but for it to have meaning the batch size must be less than 50% of the expected rows to be processed.
• This process can be adapted to implement a “stop-retry” logic so already processed rows can be skipped if you decide to cancel the execution.
• It also supports multi-statement processes (in fact, this is the real-world use of this approach) and you can achieve this with a “control” table having all the records to work with and update accordingly.
• If you want to implement an execution log you can achieve this by adding PRINT statements. Just now that this could slow down some processes, especially for very small batch sizes.

Last Update: 2018-08-23

About the author