Well, this post has been due for quite some time. I had opened up a survey/opinion poll on https://thesqldude.com/2012/07/10/user-opinion-poll-what-topics-would-you-like-to-see-posted-on-thesqldude-weblog/ sometime back and as expected most of you out there want to see more on performance tuning and query optimization.
Figure 1: Opinion Poll Results
Without much ado and by popular demand here goes a simple scenario on query tuning.
One of my engineers came to me with the following scenario…
Problem Statement
We have this Update query that is taking much longer to run compared to before. No schema changes have been made though its quite possible data changes have been happening. The update query goes like this,
update msgboard set mb_status =10, mb_changed_by ='SATCOM'
from msgboard
where mb_type = 210
Table Schema
I was easily able to reproduce this issue on my instance, here is how my table setup looks.
/****** Object: Table [dbo].[msgboard] Script Date: 12/20/2012 01:52:07 ******/
SET ANSI_NULLS ON
GO
SET QUOTED_IDENTIFIER ON
GO
SET ANSI_PADDING ON
GO
CREATE TABLE [dbo].[msgboard](
[mb_id] [int] NOT NULL,
[mb_type] [int] NOT NULL,
[mb_status] [int] NOT NULL,
[mb_changed_by] [varchar](50) NULL,
[mb_changed_dt] AS (getdate())
) ON [PRIMARY]
GO
SET ANSI_PADDING OFF
GO
CREATE UNIQUE CLUSTERED INDEX [pk_msgboard] ON [dbo].[msgboard]
(
[mb_id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON, FILLFACTOR = 100) ON [PRIMARY]
GO
CREATE NONCLUSTERED INDEX [IX_msgboard_type] ON [dbo].[msgboard]
(
[mb_type] ASC,
[mb_status] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
I populated this table with some sample data of 1000 rows using the following script.
declare @counter int
set @counter=0
declare @mbid int
set @mbid=1
declare @mbtype int
set @mbtype=200
declare @mbstatus int
set @mbstatus=10
while(@counter<1000)
begin
INSERT INTO [dbo].[msgboard]
([mb_id]
,[mb_type]
,[mb_status]
,[mb_changed_by])
VALUES
(@mbid,@mbtype,@mbstatus,'THESQLDUDE')
set @mbid+=1
set @counter+=1
end
GO
INSERT INTO [test].[dbo].[msgboard]
([mb_id]
,[mb_type]
,[mb_status]
,[mb_changed_by])
VALUES
(1001,210,10,'MANWITHNONAME')
So, I have 2 indexes on this table:-
1) A unique Clustered Index pk_msgboard on the mb_id column
2) A composite non-clustered index IX_msgboard_type on the columns mb_type and mb_status (in this order).
Query Execution Details
Here is what I observed when I ran the following update statement.
update msgboard set mb_status =10, mb_changed_by ='SATCOM'
from msgboard
where mb_type = 210
To actual execution plan looks like this,
Total Cost: 0.03124485
As you can see, this query is using the Clustered Index pk_msgboard and its performing a Scan. At this point, two questions should pop into your head
1. Why is the query using the clustered index?
2. Why is it performing a Scan operation?
Answer to question #1
Let’s look at the update query carefully. The query is updating the columns mb_status and mb_changed_by, and there is a filter on column mb_type. You might think, well, I have an index IX_msgboard_type on the mb_type column, why is SQL Server not using this non-clustered index?
To prove I was smarter than SQL Server optimizer, I forced it to use the non-clustered index by adding an index hint. The query looks like this,
update msgboard set mb_status =10, mb_changed_by ='SATCOM'
from msgboard WITH(INDEX(IX_msgboard_type))
where mb_type = 210
Total Cost: 0.03651106
- Notice that the Cost for the query with index hint is higher than the one without! If the table had more rows, we would have seen a bigger difference in cost
.
- Additionally, a Table Spool operator is now part of the execution plan. The table spool operator caches the row to be updated (In tempdb) and then feeds it to the Clustered Index update operator.
- This proves that SQL Server optimizer was actually choosing the cheapest and best plan it could come up with.
The reason SQL Server picked the clustered index is because the query is updating the columns mb_status =10 and mb_changed_by. This column mb_status is not the leading column in the NC index IX_msgboard_type, but the 2nd column. Why does this matter? It does because, the statistics for this NC index would be built on the leading column which is mb_type. Here is how the histogram looks.
dbcc show_statistics ('dbo.msgboard','IX_msgboard_type')
To avoid a bookmark lookup (Key Lookup operator from SQL 2005 onwards), SQL Server decided to use the clustered index since it covers all the columns in the table. If SQL Server were to use the NC index, it would be do a lookup or cache the results (ahem, Table Spool operator!) and then find the matching row in the clustered index to actually update the 2 columns. This is an additional cost, and this is why SQL decided not to go with this plan.
Coming to the 2nd question of why a Scan? It might surprise some of you, but a Seek does not imply good performance always. My favourite answer – “It depends”. In this case, the table only had 1001 rows and of which only 1 row (mb_id=1001) qualified for the update. When using the CI Scan, SQL Server was applying the filter on mb_type=210 and that returned 1 row back. A Clustered Index Seek in this case would not make any positive difference in query execution time.
OBJECT:([test].[dbo].[msgboard].[pk_msgboard]), WHERE:([test].[dbo].[msgboard].[mb_type]=(210)) ORDERED
To test this, I added the FORCESEEK table hint to this update query and as expected the query did not run faster.
update msgboard set mb_status =10, mb_changed_by ='SATCOM'
from msgboard WITH (FORCESEEK) where mb_type = 210
In fact it has the same cost as the query with index hint –> 0.03651106
Solution
Instead of trying to second guess the query optimizer and trying to tweak the execution plan it is producing, this issue needs a different approach –> back to basics. Based on the query design, if the column being updated mb_status was “covered” then it would help. In this case that was also also take care of, since IX_msgboard_type also includes the column mb_status, though as a non-key column. This is an important thing and this was why the “included columns” feature was introduced.
So, I modified the index IX_msgboard_type to remove the column mb_status. Instead, I added the column back as an included column
DROP INDEX [IX_msgboard_type] ON [dbo].[msgboard] WITH ( ONLINE = OFF )
GO
CREATE NONCLUSTERED INDEX [IX_msgboard_type] ON [dbo].[msgboard]
(
[mb_type] ASC
)
INCLUDE ( [mb_status]) WITH (STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
GO
The biggest advantage of having an included column in a non-clustered index is that, the query optimizer can locate all the column values within the index. The base table or clustered index data is not accessed resulting in fewer disk I/O operations and hence faster query execution. Also, the index size is lesser since non-key columns are only added to the leaf level. Many advantages indeed.
Here is how the update query plan looked after added mb_status as an included column.
My oh my, this plan looks so simple, clean and yeah faster. The actual cost is now 0.0132843, which is a big improvement over the original plan of 0.03651106
So we have achieved our original goals of
a) Getting SQL Server to use the Non-clustered index IX_msgboard_type, instead of the clustered index.
b) Getting an Seek operation instead of a Scan.
c) Query cost is cheaper.
So, hope everyone understood how beneficial included columns are on non-clustered indexes. For more information on guidelines and restriction with included columns refer this MSDN article.
That’s all for now folks. Keep tuned to this site for more.
-TheSQLDude (Sudarshan)
The SQL Dude! 