Slow query when using @variable in Where clause [modified]
-
Mika Wendelius wrote:
And I suspect that this was the way both query versions behaved?
It only happened for @c = 1
Mika Wendelius wrote:
Is that something we can rely on?
Probably, but I think it would be better to write a query independent of Date values. Frankly, I think my query is not much complicated and everybody now and then my find himself dealing with such queries. It's really sad that SQL Server is not able to handle it. Thanks for trying to help me Mika, please let me know if you come up with any solution.
What about the parameter question I asked? I was thinking about a solution where you would define a starting point as float. Initial starting point is 0 and you pass how many rows you want per page (for example 10). Next time you call this query, you would specify the starting point as the float found on the last row of the page you got and then again how many rows you want. So what this means in action is that first you create a new column:
ALTER TABLE News
ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, Date))
PERSISTEDThen you index it:
CREATE NONCLUSTERED INDEX X_StartPoint ON News (StartPoint)
Now the query would be like:
DECLARE @Start float, @Count INT
SET @Start = 0
SET @Count = 5
DECLARE @c TINYINT
SET @c = 1;
SELECT TOP(@Count)
id, title, description, StartPoint
FROM News a
WHERE Culture = @c
AND StartPoint > @Start
AND StartPoint < ((@c + 1) * 100000000)
ORDER BY StartThe need to optimize rises from a bad design.My articles[^]
-
Rob Philpott wrote:
Have you tried inserting @c into a single line temporary table and joining on that?
Dear Rob, I used the following query but it didn't make any difference:
DECLARE @Start INT, @Count INT
SET @Start = 10
SET @Count = 5DECLARE @c TINYINT
SET @c = 0;WITH paging AS (
SELECT id, title, [description], ROW_NUMBER() OVER (ORDER BY Date DESC) rownum
FROM news
JOIN (SELECT @C AS Cult) AS CC ON news.Culture = CC.Cult
--WHERE Culture = @c
)
SELECT * FROM paging WHERE rownum BETWEEN @Start AND (@Start + @Count - 1)modified on Wednesday, January 28, 2009 3:03 AM
-
What about the parameter question I asked? I was thinking about a solution where you would define a starting point as float. Initial starting point is 0 and you pass how many rows you want per page (for example 10). Next time you call this query, you would specify the starting point as the float found on the last row of the page you got and then again how many rows you want. So what this means in action is that first you create a new column:
ALTER TABLE News
ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, Date))
PERSISTEDThen you index it:
CREATE NONCLUSTERED INDEX X_StartPoint ON News (StartPoint)
Now the query would be like:
DECLARE @Start float, @Count INT
SET @Start = 0
SET @Count = 5
DECLARE @c TINYINT
SET @c = 1;
SELECT TOP(@Count)
id, title, description, StartPoint
FROM News a
WHERE Culture = @c
AND StartPoint > @Start
AND StartPoint < ((@c + 1) * 100000000)
ORDER BY StartThe need to optimize rises from a bad design.My articles[^]
Thanks Mika for the solution, But I can not make the column persisted because of the following error:
Computed column 'StartPoint' in table 'News' cannot be persisted because the column is non-deterministic. -
Thanks Mika for the solution, But I can not make the column persisted because of the following error:
Computed column 'StartPoint' in table 'News' cannot be persisted because the column is non-deterministic.Sorry, at least one typo. The date column must be in angle brackets. I made succesfully a test like this. Is it similar to your structure:
CREATE TABLE News (
[Date] datetime,
[Id] int,
[Culture] tinyint
)ALTER TABLE News
ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, [Date]))
PERSISTEDThe need to optimize rises from a bad design.My articles[^]
-
Sorry, at least one typo. The date column must be in angle brackets. I made succesfully a test like this. Is it similar to your structure:
CREATE TABLE News (
[Date] datetime,
[Id] int,
[Culture] tinyint
)ALTER TABLE News
ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, [Date]))
PERSISTEDThe need to optimize rises from a bad design.My articles[^]
Mika Wendelius wrote:
CREATE TABLE News ( [Date] datetime, [Id] int, [Culture] tinyint ) -- ALTER TABLE News ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, [Date])) PERSISTED
I can execute this query too, but when I want to alter the current News table, it tells me that the columns is non-deterministic. :confused: Apart from that, I didn't get the concept of multiplying Culture by 100000000. Why didn't you write something just like this?
ADD StartPoint AS CONVERT(float, [Date])
And yet another question, what't the difference between sorting on a float column rather than date column?
-
Mika Wendelius wrote:
CREATE TABLE News ( [Date] datetime, [Id] int, [Culture] tinyint ) -- ALTER TABLE News ADD StartPoint AS ((Culture * 100000000) + CONVERT(float, [Date])) PERSISTED
I can execute this query too, but when I want to alter the current News table, it tells me that the columns is non-deterministic. :confused: Apart from that, I didn't get the concept of multiplying Culture by 100000000. Why didn't you write something just like this?
ADD StartPoint AS CONVERT(float, [Date])
And yet another question, what't the difference between sorting on a float column rather than date column?
Maysam Mahfouzi wrote:
when I want to alter the current News table, it tells me that the columns is non-deterministic
Could you post the script for the News table.
Maysam Mahfouzi wrote:
I didn't get the concept of multiplying Culture by 100000000
The idea is to create a single column which orders the culture and date columns together. To combine those columns to a single column and to have unique values I converted the date to float so that I can add culture to it. However since Cultures are 1,2,3... I cannot add them to the date since it would change the "date portion" of the float so I decided to multiply the culture with a number big enough so that it won't get mixed with the date. For example if I would add a new record to the new right now with culture 1 the result would be 100039839.592728. After that I can use these numbers as "pointers" to rows to define a specific starting point to the query. This eliminates the need to read previous records. So what I'm after is that if you have selected news page by page, the next query can always start from the exact point where the last record came from.
The need to optimize rises from a bad design.My articles[^]
-
Maysam Mahfouzi wrote:
when I want to alter the current News table, it tells me that the columns is non-deterministic
Could you post the script for the News table.
Maysam Mahfouzi wrote:
I didn't get the concept of multiplying Culture by 100000000
The idea is to create a single column which orders the culture and date columns together. To combine those columns to a single column and to have unique values I converted the date to float so that I can add culture to it. However since Cultures are 1,2,3... I cannot add them to the date since it would change the "date portion" of the float so I decided to multiply the culture with a number big enough so that it won't get mixed with the date. For example if I would add a new record to the new right now with culture 1 the result would be 100039839.592728. After that I can use these numbers as "pointers" to rows to define a specific starting point to the query. This eliminates the need to read previous records. So what I'm after is that if you have selected news page by page, the next query can always start from the exact point where the last record came from.
The need to optimize rises from a bad design.My articles[^]
So I suppose that you are suggesting an alternative way to replace ROW_NUMBER() function, so that we do not need to calculate row number each time from scratch, am I right? If we use:
ROW_NUMBER() OVER(ORDER BY news.[Date] DESC)
each time all rows should be sorted based on Date and then row number will be assigned to them. But with your approach, we do not need to sort Date column at all. If I'm right, can we conclude that ROW_NUMBER() OVER(ORDER BY news.[Date] DESC) is not a good option for tables containing a millions of rows? here is my table script:
CREATE TABLE [News](
[ID] [bigint] IDENTITY(1,1) NOT NULL,
[Url] [varchar](300) NULL,
[Title] [nvarchar](200) NULL,
[Description] [nvarchar](3000) NULL,
[Date] [datetime] NULL,
[Rank] [float] NULL,
[ViewCount] [int] NULL,
[Vote] [smallint] NULL,
[Culture] [tinyint] NULL,
CONSTRAINT [PK_NewsItem_1] 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] -
So I suppose that you are suggesting an alternative way to replace ROW_NUMBER() function, so that we do not need to calculate row number each time from scratch, am I right? If we use:
ROW_NUMBER() OVER(ORDER BY news.[Date] DESC)
each time all rows should be sorted based on Date and then row number will be assigned to them. But with your approach, we do not need to sort Date column at all. If I'm right, can we conclude that ROW_NUMBER() OVER(ORDER BY news.[Date] DESC) is not a good option for tables containing a millions of rows? here is my table script:
CREATE TABLE [News](
[ID] [bigint] IDENTITY(1,1) NOT NULL,
[Url] [varchar](300) NULL,
[Title] [nvarchar](200) NULL,
[Description] [nvarchar](3000) NULL,
[Date] [datetime] NULL,
[Rank] [float] NULL,
[ViewCount] [int] NULL,
[Vote] [smallint] NULL,
[Culture] [tinyint] NULL,
CONSTRAINT [PK_NewsItem_1] 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]Yes you interpretaion was correct. That's what I tried to explain (but with lots of useless word between :)) It seems that SQL Server 2005 won't create this column while 2008 will. Too bad. However, I think we could do this with the datetime also
DECLARE @Start datetime, @Count INT
SET @Start = GETDATE()
SET @Count = 5
DECLARE @c TINYINT
SET @c = 1;
SELECT TOP(@Count)
id, title, description, Date
FROM News a
WHERE Culture = @c
AND Date < @Start
ORDER BY Date DESCThe idea is the same. Define a starting point to the current datetime, fetch page and when you want the next page, define the starting point to the last date in the previous page (from the last row). The date column must be unique in this version so two different news must have at least 1 ms difference in date. You can try it by testing different values for the @Start variable.
The need to optimize rises from a bad design.My articles[^]
-
Yes you interpretaion was correct. That's what I tried to explain (but with lots of useless word between :)) It seems that SQL Server 2005 won't create this column while 2008 will. Too bad. However, I think we could do this with the datetime also
DECLARE @Start datetime, @Count INT
SET @Start = GETDATE()
SET @Count = 5
DECLARE @c TINYINT
SET @c = 1;
SELECT TOP(@Count)
id, title, description, Date
FROM News a
WHERE Culture = @c
AND Date < @Start
ORDER BY Date DESCThe idea is the same. Define a starting point to the current datetime, fetch page and when you want the next page, define the starting point to the last date in the previous page (from the last row). The date column must be unique in this version so two different news must have at least 1 ms difference in date. You can try it by testing different values for the @Start variable.
The need to optimize rises from a bad design.My articles[^]
But I guess there is a problem with this approach. That is, when we show pagination buttons at the bottom of page, how can user go to page 7 from page 1 without knowing the last date of page 6?
-
But I guess there is a problem with this approach. That is, when we show pagination buttons at the bottom of page, how can user go to page 7 from page 1 without knowing the last date of page 6?
Okay, if it's possible to skip pages, then this won't work. I started to think some kind of pre-pagination, but I think that won't work either... Hmm, are the news deleted often? Could it be acceptable that if a record is deleted, a time consuming operation is executed? What I'm thinking is that if we would have a ordinal number on the rows (1,2,3,4...) and we know the maximum at every moment then the page 6 (with 10 news on page) would be from (maximum - (5 * 10)) to (maximum - (6 * 10)) + 1. Now if we create an unique index on the ordinal column we get the maximum very fast and the query will also be very fast. The downside is that if a record is deleted, the ordinals must be rearranged starting from the deleted position to the maximum. However this would be one simple operation (although it may take awhile). Do you see problems in that solution?
The need to optimize rises from a bad design.My articles[^]
-
Okay, if it's possible to skip pages, then this won't work. I started to think some kind of pre-pagination, but I think that won't work either... Hmm, are the news deleted often? Could it be acceptable that if a record is deleted, a time consuming operation is executed? What I'm thinking is that if we would have a ordinal number on the rows (1,2,3,4...) and we know the maximum at every moment then the page 6 (with 10 news on page) would be from (maximum - (5 * 10)) to (maximum - (6 * 10)) + 1. Now if we create an unique index on the ordinal column we get the maximum very fast and the query will also be very fast. The downside is that if a record is deleted, the ordinals must be rearranged starting from the deleted position to the maximum. However this would be one simple operation (although it may take awhile). Do you see problems in that solution?
The need to optimize rises from a bad design.My articles[^]
Mika Wendelius wrote:
Do you see problems in that solution?
I don't see any problem with it dude, it seems something to work :) But the question is: Is this just MY problem with pagination, or everybody else is dealing with such problem for large number of rows. Have you ever had the same problem with paging? You know my conclusion is that, maybe, pagination has never been meant to be used for large number of rows. If a user wants to browse news, it's very unlikely that he/she wants to go to page 3000, right? We can provide user with a filter on Date of news, and then he/she will be able to first filter them based on date, and then browse them by page.
-
Mika Wendelius wrote:
Do you see problems in that solution?
I don't see any problem with it dude, it seems something to work :) But the question is: Is this just MY problem with pagination, or everybody else is dealing with such problem for large number of rows. Have you ever had the same problem with paging? You know my conclusion is that, maybe, pagination has never been meant to be used for large number of rows. If a user wants to browse news, it's very unlikely that he/she wants to go to page 3000, right? We can provide user with a filter on Date of news, and then he/she will be able to first filter them based on date, and then browse them by page.
Maysam Mahfouzi wrote:
Is this just MY problem with pagination, or everybody else is dealing with such problem for large number of rows
I think I'm not the best person to answer that. The applications I mostly design act a little bit differently. Although they handle larege amount of rows I try to keep the result sets small and avoid paging at all. I think that there propably is several different solutions and the one I suggested may be one of those. Another approach could be that you page per date as you mentioned. One technical modification could be to use partitioning (but that requires Enterprise Edition) etc.
Maysam Mahfouzi wrote:
You know my conclusion is that, maybe, pagination has never been meant to be used for large number of rows
That's my understanding too. It really doesn't make sense if you're on page 212347 :) Also one thing that we haven't spoken is caching. The page number change quite seldomly so you could cache the pages (or let's say first 100 pages) and use them from cache. In that case you just need to now when news are added in order to redresh the cache. That can be done using polling or better yet using SqlDependency. Still if cache is used and you go to the data beyond cache the search mechanism should be efficient. So if you try the last suggestion and let's see what it does. Also try utilizing cache. It greatly helps you to maintain overall performance of the system.
The need to optimize rises from a bad design.My articles[^]
-
Maysam Mahfouzi wrote:
Is this just MY problem with pagination, or everybody else is dealing with such problem for large number of rows
I think I'm not the best person to answer that. The applications I mostly design act a little bit differently. Although they handle larege amount of rows I try to keep the result sets small and avoid paging at all. I think that there propably is several different solutions and the one I suggested may be one of those. Another approach could be that you page per date as you mentioned. One technical modification could be to use partitioning (but that requires Enterprise Edition) etc.
Maysam Mahfouzi wrote:
You know my conclusion is that, maybe, pagination has never been meant to be used for large number of rows
That's my understanding too. It really doesn't make sense if you're on page 212347 :) Also one thing that we haven't spoken is caching. The page number change quite seldomly so you could cache the pages (or let's say first 100 pages) and use them from cache. In that case you just need to now when news are added in order to redresh the cache. That can be done using polling or better yet using SqlDependency. Still if cache is used and you go to the data beyond cache the search mechanism should be efficient. So if you try the last suggestion and let's see what it does. Also try utilizing cache. It greatly helps you to maintain overall performance of the system.
The need to optimize rises from a bad design.My articles[^]
I'm sorry to bother you with this long conversation but I'd like to continue it till I feel like I have no more questions :) I'm thinking of a problem with last approach. That's right that it does not make sense to be on page 243546, but, think of a weblog, I can see paging for blog posts, and I want to go to last page! So, there have to be some way for me to get to last page (to see the first blog post), which may be on page 23453! The problem with your last solution is that it will not work if we want to give paging to a result set from more than one table! right?
Mika Wendelius wrote:
One technical modification could be to use partitioning (but that requires Enterprise Edition) etc.
How can partitioning help? Can you show me a starting point? Think of a site like friendfeed, in this site I can see recent activities of a user. These activities have been probably collected from different tables in database. In this scenario, what happens if I want to see the very first activities of user? (I have to go to last page of activities)
-
I'm sorry to bother you with this long conversation but I'd like to continue it till I feel like I have no more questions :) I'm thinking of a problem with last approach. That's right that it does not make sense to be on page 243546, but, think of a weblog, I can see paging for blog posts, and I want to go to last page! So, there have to be some way for me to get to last page (to see the first blog post), which may be on page 23453! The problem with your last solution is that it will not work if we want to give paging to a result set from more than one table! right?
Mika Wendelius wrote:
One technical modification could be to use partitioning (but that requires Enterprise Edition) etc.
How can partitioning help? Can you show me a starting point? Think of a site like friendfeed, in this site I can see recent activities of a user. These activities have been probably collected from different tables in database. In this scenario, what happens if I want to see the very first activities of user? (I have to go to last page of activities)
Maysam Mahfouzi wrote:
I'm sorry to bother you with this long conversation
You're not bothering me at all :)
Maysam Mahfouzi wrote:
The problem with your last solution is that it will not work if we want to give paging to a result set from more than one table
That's correct since the numbers are precalculated per table.
Maysam Mahfouzi wrote:
How can partitioning help
Since partitioning divides the rows into smaller sets, it improves performance and makes parallel querying easier for the dmbs. To get a quick look at partitioning, see: Partitioned Table and Index Concepts[^].
Maysam Mahfouzi wrote:
These activities have been probably collected from different tables in database
This is the same problem as multiple tables in a query. I don't have any good idea immediately how to do it using SQL but again a technical tool could be using indexed views[^] (again an Enterprise Edition feature). It could be something like that you do the join for all rows and include row numbering into the result set. After that the view is indexed (this partly solves all multi table queries and ordering).
The need to optimize rises from a bad design.My articles[^]
-
Maysam Mahfouzi wrote:
I'm sorry to bother you with this long conversation
You're not bothering me at all :)
Maysam Mahfouzi wrote:
The problem with your last solution is that it will not work if we want to give paging to a result set from more than one table
That's correct since the numbers are precalculated per table.
Maysam Mahfouzi wrote:
How can partitioning help
Since partitioning divides the rows into smaller sets, it improves performance and makes parallel querying easier for the dmbs. To get a quick look at partitioning, see: Partitioned Table and Index Concepts[^].
Maysam Mahfouzi wrote:
These activities have been probably collected from different tables in database
This is the same problem as multiple tables in a query. I don't have any good idea immediately how to do it using SQL but again a technical tool could be using indexed views[^] (again an Enterprise Edition feature). It could be something like that you do the join for all rows and include row numbering into the result set. After that the view is indexed (this partly solves all multi table queries and ordering).
The need to optimize rises from a bad design.My articles[^]