Slow deletion of records when a trigger is enabled

The row-versioning framework introduced in SQL Server 2005 is used to support a number of features, including the new transaction isolation levels READ_COMMITTED_SNAPSHOT and SNAPSHOT. Even when neither of these isolation levels are enabled, row-versioning is still used for AFTER triggers (to facilitate generation of the inserted and deleted pseudo-tables), MARS, and (in a separate version store) online indexing.

As documented, the engine may add a 14-byte postfix to each row of a table that is versioned for any of these purposes. This behaviour is relatively well-known, as is the addition of the 14-byte data to every row of an index that is rebuilt online with a row-versioning isolation level enabled. Even where the isolation levels are not enabled, one extra byte is added to non-clustered indexes only when rebuilt ONLINE.

Where an AFTER trigger is present, and versioning would otherwise add 14 bytes per row, an optimization exists within the engine to avoid this, but where a ROW_OVERFLOW or LOB allocation cannot occur. In practice, this means the maximum possible size of a row must be less than 8060 bytes. In calculating maximum possible row sizes, the engine assumes for example that a VARCHAR(460) column could contain 460 characters.

The behaviour is easiest to see with an AFTER UPDATE trigger, though the same principle applies to AFTER DELETE. The following script creates a table with a maximum in-row length of 8060 bytes. The data fits on a single page, with 13 bytes of free space on that page. A no-op trigger exists, so the page is split and versioning information added:

USE Sandpit;
GO
CREATE TABLE dbo.Example
(
    ID          integer NOT NULL IDENTITY(1,1),
    Value       integer NOT NULL,
    Padding1    char(42) NULL,
    Padding2    varchar(8000) NULL,

    CONSTRAINT PK_Example_ID
    PRIMARY KEY CLUSTERED (ID)
);
GO
WITH
    N1 AS (SELECT 1 AS n UNION ALL SELECT 1),
    N2 AS (SELECT L.n FROM N1 AS L CROSS JOIN N1 AS R),
    N3 AS (SELECT L.n FROM N2 AS L CROSS JOIN N2 AS R),
    N4 AS (SELECT L.n FROM N3 AS L CROSS JOIN N3 AS R)
INSERT TOP (137) dbo.Example
    (Value)
SELECT
    ROW_NUMBER() OVER (ORDER BY (SELECT 0))
FROM N4;
GO
ALTER INDEX PK_Example_ID 
ON dbo.Example 
REBUILD WITH (FILLFACTOR = 100);
GO
SELECT
    ddips.index_type_desc,
    ddips.alloc_unit_type_desc,
    ddips.index_level,
    ddips.page_count,
    ddips.record_count,
    ddips.max_record_size_in_bytes
FROM sys.dm_db_index_physical_stats(DB_ID(), OBJECT_ID(N'dbo.Example', N'U'), 1, 1, 'DETAILED') AS ddips
WHERE
    ddips.index_level = 0;
GO
CREATE TRIGGER ExampleTrigger
ON dbo.Example
AFTER DELETE, UPDATE
AS RETURN;
GO
UPDATE dbo.Example
SET Value = -Value
WHERE ID = 1;
GO
SELECT
    ddips.index_type_desc,
    ddips.alloc_unit_type_desc,
    ddips.index_level,
    ddips.page_count,
    ddips.record_count,
    ddips.max_record_size_in_bytes
FROM sys.dm_db_index_physical_stats(DB_ID(), OBJECT_ID(N'dbo.Example', N'U'), 1, 1, 'DETAILED') AS ddips
WHERE
    ddips.index_level = 0;
GO
DROP TABLE dbo.Example;

The script produces the output shown below. The single-page table is split into two pages, and the maximum physical row length has increased from 57 to 71 bytes (= +14 bytes for the row-versioning information).

DBCC PAGE shows that the single updated row has Record Attributes = NULL_BITMAP VERSIONING_INFO Record Size = 71, whereas all other rows in the table have Record Attributes = NULL_BITMAP; record Size = 57.

The same script, with the UPDATE replaced by a single row DELETE produces the output shown:

DELETE dbo.Example
WHERE ID = 1;

There is one fewer row in total (of course!), but the maximum physical row size has not increased. Row versioning information is only added to rows needed for the trigger pseudo-tables, and that row was ultimately deleted. The page split remains, however. This page-splitting activity is responsible for the slow performance observed when the trigger was present. If the definition of the Padding2 column is changed from varchar(8000) to varchar(7999), the page no longer splits.

Also see this blog post by SQL Server MVP Dmitri Korotkevitch, which also discusses the impact on fragmentation.

Well here is the official response from Microsoft...which I think is a major design flaw.

11/14/2011 - Official response has changed. They are not using the transaction log as previously stated. The are using the internal store (row level) to copy the changed data into. They still can't determine why it's taken so long.

We decided to use Instead Of triggers in lieu of after delete triggers.

The AFTER part of the trigger causes us to have to read through the transaction log after the deletes complete and build the trigger inserted/deleted table. This is where we spend the vast amount of time and is by design for the AFTER part of the trigger. INSTEAD OF trigger would prevent this behavior of scanning the transaction log and building an inserted/deleted table. Also, as it was observed that things are much faster if we drop all columns with nvarchar(max), which makes sense due to the fact that it is considered LOB data. Please have alook at below article for more informaiton regarding In-Row data:

http://msdn.microsoft.com/en-us/library/ms189087.aspx

Summary: AFTER trigger requires scanning back through the transaction log after the delete finishes then we have to build and inserted/deleted table which requires more usage of the transaction log and time.

So as an action plan, this is what we suggest at this time:

A) Limit the number of rows deleted in each transaction or
B) Increase timeout settings or
C) Don't use AFTER trigger or trigger at all or
D) Limit usage of nvarchar(max) datatypes.

According to the plan everything is going correctly. You can try writing the delete as a JOIN instead of an IN which will give you a different plan.

DELETE m
FROM MAIN m
JOIN Secondary s ON m.ID = s.ValueInt1
AND s.SetTMGUID = '9DDD2C8DD3864EA7B78DA22B2ED572D7'

I'm not sure how much that will help however. When the delete is running with the triggers on the table what is the wait type for the session doing the delete?