• 3/11/2015

Using the TOP option with modifications

T-SQL supports using the TOP filter with modification statements. This section first describes this capability, and then its limitation and a workaround for the limitation. Then it describes a practical use case for this capability when you need to delete a large number of rows.

In my examples, I’ll use a table called MyOrders. Run the following code to create this table as an initial copy of the Orders table in the PerformanceV3 database:

USE PerformanceV3;
IF OBJECT_ID(N'dbo.MyOrders', N'U') IS NOT NULL DROP TABLE dbo.MyOrders;
SELECT * INTO dbo.MyOrders FROM dbo.Orders;
CREATE UNIQUE CLUSTERED INDEX idx_od_oid ON dbo.MyOrders(orderdate, orderid);

TOP with modifications

With T-SQL, you can use TOP with modification statements. Those statements are INSERT TOP, DELETE TOP, UPDATE TOP, and MERGE TOP. This means the statement will stop modifying rows once the requested number of rows are affected. For example, the following statement deletes 50 rows from the table MyOrders:

DELETE TOP (50) FROM dbo.MyOrders;

When you use TOP in a SELECT statement, you can control which rows get chosen using the ORDER BY clause. But modification statements don’t have an ORDER BY clause. This means you can indicate how many rows you want to modify, but not based on what order—at least, not directly. So the preceding statement deletes 50 rows, but you cannot control which 50 rows get deleted. You should consider the order as being arbitrary. In practice, it depends on optimization and data layout.

This limitation is a result of the design choice that the TOP ordering is to be defined by the traditional ORDER BY clause. The traditional ORDER BY clause was originally designed to define presentation order, and it is available only to the SELECT statement. Had the design of TOP been different, with its own ordering specification that is not related to presentation ordering, it would have been natural to use also with modification statements. Here’s an example for what such a design might have looked like (but don’t run the code, because this syntax isn’t supported):

DELETE TOP (50) OVER(ORDER BY orderdate, orderid) FROM dbo.MyOrders;

Fortunately, when you do need to control which rows get chosen, you can use a simple trick as a workaround. Use a SELECT query with a TOP filter and an ORDER BY clause. Define a table expression based on that query. Then issue the modification against the table expression, like so:

  SELECT TOP (50) *
  FROM dbo.MyOrders
  ORDER BY orderdate, orderid

In practice, the rows from the underlying table will be affected. You can think of the modification as being defined through the table expression.

The OFFSET-FETCH filter is not supported directly with modification statements, but you can use a similar trick like the one with TOP.

Modifying in chunks

Having TOP supported by modification statements without the ability to indicate order might seem futile, but there is a practical use case involving modifying large volumes of data. As an example, suppose you need to delete all rows from the MyOrders table where the order date is before 2013. The table in our example is fairly small, having about 1,000,000 rows. But imagine there were 100,000,000 rows, and the number of rows to delete was about 50,000,000. If the table was partitioned (say, by year), things would be easy and highly efficient. You switch a partition out to a staging table and then drop the staging table. However, what if the table is currently not partitioned? The intuitive thing to do is to issue a simple DELETE statement to do the job as a single transaction, like so (but don’t run this statement):

DELETE FROM dbo.MyOrders WHERE orderdate < '20130101';

Such an approach can get you into trouble in a number of ways.

A DELETE statement is fully logged, unlike DROP TABLE and TRUNCATE TABLE statements. Log writes are sequential; therefore, log-intensive operations tend to be slow and hard to optimize beyond a certain point. For example, deleting 50,000,000 rows can take many minutes to finish.

There’s a section in the log considered to be the active portion, starting with the oldest open transaction and ending with the current pointer in the log. The active portion cannot be recycled. So when you have a long-running transaction, it can cause the transaction log to expand, sometimes well beyond its typical size for your database. This can be an issue if you have limited disk space.

Modification statements acquire exclusive locks on the modified resources (row or page locks, as decided by SQL Server dynamically), and exclusive locks are held until the transaction finishes. Each lock is represented by a memory structure that is approximately 100 bytes in size. Acquiring a large number of locks has two main drawbacks. For one, it requires large amounts of memory. Second, it takes time to allocate the memory structures, which adds to the time it takes the transaction to complete. To reduce the memory footprint and allow a faster process, SQL Server will attempt to escalate from the initial granularity of locks (row or page) to a table lock (or partition, if configured). The first trigger for SQL Server to attempt escalation is when the same transaction reaches 5,000 locks against the same object. If unsuccessful (for example, another transaction is holding locks that the escalated lock would be in conflict with), SQL Server will keep trying to escalate every additional 1,250 locks. When escalation succeeds, the transaction locks the entire table (or partition) until it finishes. This behavior can cause concurrency problems.

If you try to terminate such a large modification that is in progress, you will face the consequences of a rollback. If the transaction was already running for a while, it will take a while for the rollback to finish—typically, more than the original work.

To avoid the aforementioned problems, the recommended approach to apply a large modification is to do it in chunks. For our purge process, you can run a loop that executes a DELETE TOP statement repeatedly until all qualifying rows are deleted. You want to make sure that the chunk size is not too small so that the process will not take forever, but you want it to be small enough not to trigger lock escalation. The tricky part is figuring out the chunk size. It takes 5,000 locks before SQL Server attempts escalation, but how does this translate to the number of rows you’re deleting? SQL Server could decide to use row or page locks initially, plus when you delete rows from a table, SQL Server deletes rows from the indexes that are defined on the table. So it’s hard to predict what the ideal number of rows is without testing.

A simple solution could be to test different numbers while running a trace or an extended events session with a lock-escalation event. For example, I ran the following extended events session with a Live Data window open, while issuing DELETE TOP statements from a session with session ID 53:

ADD EVENT sqlserver.lock_escalation(
    WHERE ([sqlserver].[session_id]=(53)));

I started with 10,000 rows using the following statement:

DELETE TOP (10000) FROM dbo.MyOrders WHERE orderdate < '20130101';

Then I adjusted the number, increasing or decreasing it depending on whether an escalation event took place or not. In my case, the first point where escalation happened was somewhere between 6,050 and 6,100 rows. Once you find it, you don’t want to use that point minus 1. For example, if you add indexes later on, the point will become lower. To be on the safe side, I take the number that I find in my testing and divide it by two. This should leave enough room for the future addition of indexes. Of course, it’s worthwhile to retest from time to time to see if the number needs to be adjusted.

Once you have the chunk size determined (say, 3,000), you implement the purge process as a loop that deletes one chunk of rows at a time using a DELETE TOP statement, like so:


WHILE 1 = 1
  DELETE TOP (3000) FROM dbo.MyOrders WHERE orderdate < '20130101';

The code uses an infinite loop. Every execution of the DELETE TOP statement deletes up to 3,000 rows and commits. As soon as the number of affected rows is lower than 3,000, you know that you’ve reached the last chunk, so the code breaks from the loop. If this process is running (and during peak hours), you want to abort it, and it’s quite safe to stop it. Only the current chunk will undergo a rollback. You can then run it again in the next window you have for this and the process will simply pick up where it left off.