September 5, 2012 at 8:23 am
SQL Kiwi (9/4/2012)
Blog post is up: http://bit.ly/ComputeScalar
As always, top notch stuff. I admire how you keep in-depth details so simple yet so effective. Not to mention there is always a new learning experience. Hats off to you 😉
September 5, 2012 at 8:34 am
Jeff Moden (9/5/2012)
I have to agree with Peter, here. If I extrapolate some of the things you said in your fine blog post, I can see possible explanations for why "Divide'n'Conquer" methods, such as using Temp Tables to store interim results instead of "All-in-one" queries, might execute so very much faster. Would you agree in that area?
There are many reasons to like divide-and-conquer and explicit materialization, but I'm not sure I follow your exact line of thinking here. Feel free to ping me via email if it's something you want to chat about. 2:30am here now, so it will probably be tomorrow before I respond.
Peter, Chris, Usman: Thanks guys!
Paul White
SQLPerformance.com
SQLkiwi blog
@SQL_Kiwi
September 10, 2012 at 6:37 am
Paul's CLR splitter has two features which mean it shouldn't be compared directly to DelimitedSplit8K based splitters.
The first is that it doesn't handle the trailing delimiter correctly. Using comma as a delimiter, the string '55555,' returns one item "55555", not two "55555" and "". Using Jeff's functional test cases (17 strings) it returns 45 rows and not 52.
The second is that it doesn't use collations for the comparison of string to delimiter. If the DelimitedSplit8K based splitters follow Usman's lead and use a binary collation you get the same effect and roughly a doubling in speed. The use of collations in this case is not very helpful. It's common to have a Case Insensitive collation so you could use LATIN SMALL LETTER O as a delimater and match using LATIN CAPITAL LETTER O. Less common is to use an Accent Insensitive collation where you could match using LATIN CAPITAL LETTER O WITH DIAERESIS. I can't see any use for this behaviour besides confusing yourself and others so I recommend using a binary collation. In my testing, Usman's N1 splitter is still the fastest non-CLR splitter even when binary collation is added to the DelimitedSplit8K variants.
I also think Jeff's test harness should follow Usman's example and use local variables to dispose of the results cheaply rather than store them in temp tables. Storing the results is OK for checking that the splitter works properly but the write performance of tempdb can easily dominate the result and make it harder to detect differences. Unless of course this causes the optimiser behaviour to be different.
This has been (and continues to be) a very educational thread. Just when I thought it'd run its course it fires up with a ton of ideas about XML splitters. Thanks to all who contributed.
-mark
September 10, 2012 at 7:34 am
SQL Kiwi (9/5/2012)
Jeff Moden (9/5/2012)
I have to agree with Peter, here. If I extrapolate some of the things you said in your fine blog post, I can see possible explanations for why "Divide'n'Conquer" methods, such as using Temp Tables to store interim results instead of "All-in-one" queries, might execute so very much faster. Would you agree in that area?There are many reasons to like divide-and-conquer and explicit materialization, but I'm not sure I follow your exact line of thinking here. Feel free to ping me via email if it's something you want to chat about. 2:30am here now, so it will probably be tomorrow before I respond.
Peter, Chris, Usman: Thanks guys!
Thanks for the feedback and the offer, Paul. What I meant was that it's one possible explanation for why "Divide'n'Conquer" is frequently much faster than "All-in-one".
--Jeff Moden
Change is inevitable... Change for the better is not.
September 10, 2012 at 7:41 am
m.t.cleary (9/10/2012)
Paul's CLR splitter has two features which mean it shouldn't be compared directly to DelimitedSplit8K based splitters.The first is that it doesn't handle the trailing delimiter correctly. Using comma as a delimiter, the string '55555,' returns one item "55555", not two "55555" and "". Using Jeff's functional test cases (17 strings) it returns 45 rows and not 52.
The second is that it doesn't use collations for the comparison of string to delimiter. If the DelimitedSplit8K based splitters follow Usman's lead and use a binary collation you get the same effect and roughly a doubling in speed. The use of collations in this case is not very helpful. It's common to have a Case Insensitive collation so you could use LATIN SMALL LETTER O as a delimater and match using LATIN CAPITAL LETTER O. Less common is to use an Accent Insensitive collation where you could match using LATIN CAPITAL LETTER O WITH DIAERESIS. I can't see any use for this behaviour besides confusing yourself and others so I recommend using a binary collation. In my testing, Usman's N1 splitter is still the fastest non-CLR splitter even when binary collation is added to the DelimitedSplit8K variants.
I also think Jeff's test harness should follow Usman's example and use local variables to dispose of the results cheaply rather than store them in temp tables. Storing the results is OK for checking that the splitter works properly but the write performance of tempdb can easily dominate the result and make it harder to detect differences. Unless of course this causes the optimiser behaviour to be different.
This has been (and continues to be) a very educational thread. Just when I thought it'd run its course it fires up with a ton of ideas about XML splitters. Thanks to all who contributed.
-mark
Oddly enough, I normally use throw away variables for such testing but, since I usually normalize such data and store it in a table, it was important to me to store it in a table.
I quickly looked back at Usman's recent posts and don't see a function with "N1" in it. Would yoou do me the favor of posting the exact link for which of Usman's code you're speaking about so I don't make the mistake of looking at the wrong piece of code? Thanks. (You cann do such a thing by clicking on the post number at the bottom of the left side of the post.)
--Jeff Moden
Change is inevitable... Change for the better is not.
September 10, 2012 at 8:23 am
m.t.cleary (9/10/2012)
Paul's CLR splitter has two features which mean it shouldn't be compared directly to DelimitedSplit8K based splitters.
More than two (it supports very long strings and Unicode) but overall the comparison is a fair one.
The first is that it doesn't handle the trailing delimiter correctly. Using comma as a delimiter, the string '55555,' returns one item "55555", not two "55555" and "". Using Jeff's functional test cases (17 strings) it returns 45 rows and not 52.
I have never found that way of treating the trailing-delimiter case intuitive or useful, but it is a trivial code change if you do need it to work that way:
CREATE ASSEMBLY Split
FROM 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
WITH PERMISSION_SET = SAFE;
GO
CREATE FUNCTION dbo.SplitterB
(
@Input nvarchar(max),
@Delimiter nchar(1)
)
RETURNS TABLE
(
sequence int NULL,
item nvarchar(4000) NULL
)
AS
EXTERNAL NAME Split.UserDefinedFunctions.SplitterB;
GO
Source:
using System.Collections;
using System.Data.SqlTypes;
using Microsoft.SqlServer.Server;
public partial class UserDefinedFunctions
{
/**
* How SQL Server SQLCLR table-valued functions work:
*
* 1. SQL Server passes the input parameters in to the function and receives an enumeration object in return
* 2. SQL Server calls the MoveNext() method on the enumeration object
* 3. If the MoveNext() call returns true, SQL Server calls the Current() method to get an object for the current row
* 4. SQL Server calls the FillRow method to obtain column values for the current row
* 5. Repeat from step 2, until MoveNext() returns false
*
* */
[SqlFunction
(
DataAccess = DataAccessKind.None, // No user data access by this function
SystemDataAccess = SystemDataAccessKind.None, // No system data access by this function
IsDeterministic = true, // This function is deterministic
IsPrecise = true, // This function is precise
FillRowMethodName = "FillRow", // The method called by SQL Server to obtain the next row
TableDefinition =
"sequence INT, item NVARCHAR(4000)" // Returned table definition
)
]
// 1. SQL Server passes input parameters and receives an enumration object
public static IEnumerator SplitterB
(
[SqlFacet(MaxSize = -1)] SqlChars Input,
char Delimiter
)
{
return Input.IsNull ?
new SplitEnumerator(new char[0], char.MinValue) :
new SplitEnumerator(Input.Value, Delimiter);
}
// The enumeration object
struct SplitEnumerator : IEnumerator
{
// Constructor (called once when the object is created)
internal SplitEnumerator(char[] Input, char Delimiter)
{
// Save references
input = Input;
delimiter = Delimiter;
// Remember the length of the character array
length = input.Length;
// Structure holding split rows
record = new SplitRow();
// Starting at the first character
start = 0;
}
// Enumerator implementation
#region IEnumerator Methods
// 2. SQL Server calls the MoveNext() method on the enumeration object
bool IEnumerator.MoveNext()
{
// No more rows?
if (start > length) { return false; }
// Find the next delimiter
for (int i = start; i < length; i++)
{
if (input == delimiter)
{
// Increment the sequence number
record.Sequence++;
// Save the split element
record.Item = new string(input, start, i - start);
// Set the next element search start point
start = i + 1;
return true;
}
}
// Last item
record.Sequence++;
record.Item = new string(input, start, length - start);
start = length + 1;
return true;
}
// 3. SQL Server calls the Current() method to get an object for the current row
// (We pack the current row data in an OutputRecord structure)
object IEnumerator.Current
{
get { return record; }
}
// Required by the IEnumerator interface, but not needed for this implementation
void IEnumerator.Reset()
{
throw new System.NotImplementedException();
}
#endregion
readonly char[] input; // Reference to the string to be split
readonly int length; // Length of the input string
readonly char delimiter; // The delimiter character
int start; // Current search start position
SplitRow record; // Each row to be returned
}
// 4. SQL Server calls the FillRow method to obtain column values for the current row
public static void FillRow(object obj, out int sequence, out string item)
{
// The passed-in object is an OutputRecord
var r = (SplitRow)obj;
// Set the output parameter values
sequence = r.Sequence;
item = r.Item;
}
// Structure used to hold each row
struct SplitRow
{
internal int Sequence { get; set; } // Sequence of the element
internal string Item { get; set; } // The element
}
};
The second is that it doesn't use collations for the comparison of string to delimiter [...] I can't see any use for this behaviour besides confusing yourself and others so I recommend using a binary collation.
Yes that's why the SQLCLR uses binary comparisons. It is easy to carry over the collation settings from the caller, but I have never felt the need to do this for string splitting.
Paul White
SQLPerformance.com
SQLkiwi blog
@SQL_Kiwi
September 11, 2012 at 5:23 am
I agree that the CLR has additional and desirable features, but if we're comparing apples we need to make sure the functionality we are comparing is the same. My first test with a new variant is to verify it works properly using a full outer join against the results of a known good splitter and compare.
I also agree that it doesn't make sense in this case to use collations and I suggest we follow Usman's lead and use Latin1_General_BIN. Most characters used as delimiters will not have accented or upper/lower case version. There are other cases where it might - eg multi-character delimiters where you might be splitting on either <A or <a, or breaking on whitespace. There are probably faster ways of doing that than using collations.
I got a lot out of your post on scalar expressions - thank you.
-mark
SQL Kiwi (9/10/2012)
m.t.cleary (9/10/2012)
Paul's CLR splitter has two features which mean it shouldn't be compared directly to DelimitedSplit8K based splitters.More than two (it supports very long strings and Unicode) but overall the comparison is a fair one.
The first is that it doesn't handle the trailing delimiter correctly. Using comma as a delimiter, the string '55555,' returns one item "55555", not two "55555" and "". Using Jeff's functional test cases (17 strings) it returns 45 rows and not 52.
I have never found that way of treating the trailing-delimiter case intuitive or useful, but it is a trivial code change if you do need it to work that way:
CREATE ASSEMBLY Split
FROM 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
WITH PERMISSION_SET = SAFE;
GO
CREATE FUNCTION dbo.SplitterB
(
@Input nvarchar(max),
@Delimiter nchar(1)
)
RETURNS TABLE
(
sequence int NULL,
item nvarchar(4000) NULL
)
AS
EXTERNAL NAME Split.UserDefinedFunctions.SplitterB;
GO
Source:
using System.Collections;
using System.Data.SqlTypes;
using Microsoft.SqlServer.Server;
public partial class UserDefinedFunctions
{
/**
* How SQL Server SQLCLR table-valued functions work:
*
* 1. SQL Server passes the input parameters in to the function and receives an enumeration object in return
* 2. SQL Server calls the MoveNext() method on the enumeration object
* 3. If the MoveNext() call returns true, SQL Server calls the Current() method to get an object for the current row
* 4. SQL Server calls the FillRow method to obtain column values for the current row
* 5. Repeat from step 2, until MoveNext() returns false
*
* */
[SqlFunction
(
DataAccess = DataAccessKind.None, // No user data access by this function
SystemDataAccess = SystemDataAccessKind.None, // No system data access by this function
IsDeterministic = true, // This function is deterministic
IsPrecise = true, // This function is precise
FillRowMethodName = "FillRow", // The method called by SQL Server to obtain the next row
TableDefinition =
"sequence INT, item NVARCHAR(4000)" // Returned table definition
)
]
// 1. SQL Server passes input parameters and receives an enumration object
public static IEnumerator SplitterB
(
[SqlFacet(MaxSize = -1)] SqlChars Input,
char Delimiter
)
{
return Input.IsNull ?
new SplitEnumerator(new char[0], char.MinValue) :
new SplitEnumerator(Input.Value, Delimiter);
}
// The enumeration object
struct SplitEnumerator : IEnumerator
{
// Constructor (called once when the object is created)
internal SplitEnumerator(char[] Input, char Delimiter)
{
// Save references
input = Input;
delimiter = Delimiter;
// Remember the length of the character array
length = input.Length;
// Structure holding split rows
record = new SplitRow();
// Starting at the first character
start = 0;
}
// Enumerator implementation
#region IEnumerator Methods
// 2. SQL Server calls the MoveNext() method on the enumeration object
bool IEnumerator.MoveNext()
{
// No more rows?
if (start > length) { return false; }
// Find the next delimiter
for (int i = start; i < length; i++)
{
if (input == delimiter)
{
// Increment the sequence number
record.Sequence++;
// Save the split element
record.Item = new string(input, start, i - start);
// Set the next element search start point
start = i + 1;
return true;
}
}
// Last item
record.Sequence++;
record.Item = new string(input, start, length - start);
start = length + 1;
return true;
}
// 3. SQL Server calls the Current() method to get an object for the current row
// (We pack the current row data in an OutputRecord structure)
object IEnumerator.Current
{
get { return record; }
}
// Required by the IEnumerator interface, but not needed for this implementation
void IEnumerator.Reset()
{
throw new System.NotImplementedException();
}
#endregion
readonly char[] input; // Reference to the string to be split
readonly int length; // Length of the input string
readonly char delimiter; // The delimiter character
int start; // Current search start position
SplitRow record; // Each row to be returned
}
// 4. SQL Server calls the FillRow method to obtain column values for the current row
public static void FillRow(object obj, out int sequence, out string item)
{
// The passed-in object is an OutputRecord
var r = (SplitRow)obj;
// Set the output parameter values
sequence = r.Sequence;
item = r.Item;
}
// Structure used to hold each row
struct SplitRow
{
internal int Sequence { get; set; } // Sequence of the element
internal string Item { get; set; } // The element
}
};
The second is that it doesn't use collations for the comparison of string to delimiter [...] I can't see any use for this behaviour besides confusing yourself and others so I recommend using a binary collation.
Yes that's why the SQLCLR uses binary comparisons. It is easy to carry over the collation settings from the caller, but I have never felt the need to do this for string splitting.
September 11, 2012 at 5:33 am
m.t.cleary (9/11/2012)
I agree that the CLR has additional and desirable features, but if we're comparing apples we need to make sure the functionality we are comparing is the same. My first test with a new variant is to verify it works properly using a full outer join against the results of a known good splitter and compare.
Agreed. I was surprised no-one picked up the difference before (I guess it goes to show how many people even tried the CLR solution). By the way, I did try not to be defensive about the mickey-mouse CLR I wrote for Jeff in my last post, but I may not have succeeded in that. Apologies if I came off snarky - it wasn't intended.
Paul White
SQLPerformance.com
SQLkiwi blog
@SQL_Kiwi
September 11, 2012 at 5:34 am
The recent discussion has hit an interesting topic, namely that of collations. While I can subscribe to the idea that it is ok for the delimiter being compared by binary means, I don't like to see side effects on the function output itself when that is further processed. Having to specify collations everywhere would make the function cumbersome to use. It might even throw off the optimizer in not using indexes due to collation differences.
An example scenario could be this:
The outcome of the splitter is used to match the substrings against a table of specific words (indexed) and for each word we count the number of occurrences in the original separated input sting. Assuming both the words in that table and the original string are in the same collation, the code should be completely transparent in terms of collations, else if looses much of its appeal.
September 11, 2012 at 6:01 am
Jeff,
Usman’s N1 splitter is in http://www.sqlservercentral.com/Forums/FindPost1303895.aspx
On the subject of benchmark frameworks, I have some sympathy for storing the output. Splitters are often used where you are going insert the results and real world measurements are the ultimate test. I was playing with Usman’s splitter trying to work out where the speed differences are because the tests you’d done with physical tables seemed to indicate cte’s were faster. I found that storing a the output increased the variability and masked the small differences I was looking for.
What I learnt was:
•If you use a physical tally table, WITH (NOLOCK) makes quite a difference.
•Using COLLATE Latin1_General_BIN on the CHARINDEX makes a substantial difference.
•Using COLLATE Latin1_General_BIN on the substring() = @pDelimiter comparison helps too but not as much
•Combining the generation of cteStart and cteLen into a single Select is slightly slower than doing it in two selects
•Using a self join on cteStart to generate cteLen is BAD – cte’s are actually text macros so the self join means cteStart is generated from scratch twice
•Using a three way 20 element cross join for E4 is slightly faster than the 10 element two way joins
•Using a two way 90 element cross join is slightly faster again. (I used table value constructors instead of select 1 but I doubt it matters.)
I’ve experimented with taking the following Usman’s structure with a cte but can’t match his speed. The fastest cte splitter I got just added COLLATION Latin1_General_BIN to the DelimitedSplit8K_T1 splitter. Everything else makes things worse or no better.
I've been testing using Usman's framework combined with your functional tests (figure 22 in the article).
-mark
-mark
September 11, 2012 at 6:07 am
Paul,
I didn't take offence. I was similarly concerned I was a bit abrupt writing in a hurry late at night.
Can you avoid the conversion to Unicode by casting to a byte array?
-mark
September 11, 2012 at 6:13 am
Good point. The function should cast the output back to COLLATION DATABASE_DEFAULT. Another plus for CLR?
-mark
peter-757102 (9/11/2012)
The recent discussion has hit an interesting topic, namely that of collations. While I can subscribe to the idea that it is ok for the delimiter being compared by binary means, I don't like to see side effects on the function output itself when that is further processed. Having to specify collations everywhere would make the function cumbersome to use. It might even throw off the optimizer in not using indexes due to collation differences.An example scenario could be this:
The outcome of the splitter is used to match the substrings against a table of specific words (indexed) and for each word we count the number of occurrences in the original separated input sting. Assuming both the words in that table and the original string are in the same collation, the code should be completely transparent in terms of collations, else if looses much of its appeal.
September 11, 2012 at 8:32 am
SQL Kiwi (9/11/2012)
m.t.cleary (9/11/2012)
I agree that the CLR has additional and desirable features, but if we're comparing apples we need to make sure the functionality we are comparing is the same. My first test with a new variant is to verify it works properly using a full outer join against the results of a known good splitter and compare.Agreed. I was surprised no-one picked up the difference before (I guess it goes to show how many people even tried the CLR solution).
I did try it but only to measure the execution speed 🙂 If I would have been allowed to use it, I also might have picked up this difference as well :hehe: But seriously, it was a nice catch. Another thing I just noticed is the return data type for the item is nvarchar(4000). Should not it be nvarchar(max) ?
September 11, 2012 at 8:37 am
m.t.cleary (9/11/2012)
Jeff,Usman’s N1 splitter is in http://www.sqlservercentral.com/Forums/FindPost1303895.aspx
On the subject of benchmark frameworks, I have some sympathy for storing the output. Splitters are often used where you are going insert the results and real world measurements are the ultimate test. I was playing with Usman’s splitter trying to work out where the speed differences are because the tests you’d done with physical tables seemed to indicate cte’s were faster. I found that storing a the output increased the variability and masked the small differences I was looking for.
What I learnt was:
•If you use a physical tally table, WITH (NOLOCK) makes quite a difference.
I do not think this is the major contributor. No string concatenation (was part of old physical tally table splitter) and use of data type INT instead of BIGINT(so no implicit conversions when dealing upto varchar(8000) strings ) seems to help a lot.
September 11, 2012 at 8:41 am
m.t.cleary (9/11/2012)
peter-757102 (9/11/2012)
The recent discussion has hit an interesting topic, namely that of collations. While I can subscribe to the idea that it is ok for the delimiter being compared by binary means, I don't like to see side effects on the function output itself when that is further processed. Having to specify collations everywhere would make the function cumbersome to use. It might even throw off the optimizer in not using indexes due to collation differences.An example scenario could be this:
The outcome of the splitter is used to match the substrings against a table of specific words (indexed) and for each word we count the number of occurrences in the original separated input sting. Assuming both the words in that table and the original string are in the same collation, the code should be completely transparent in terms of collations, else if looses much of its appeal.
Good point. The function should cast the output back to COLLATION DATABASE_DEFAULT. Another plus for CLR?
-mark
I do not find the reason why? :unsure: The collation part is only for the delimiter comparison. But the item split would be of the same collation.
Viewing 15 posts - 421 through 435 (of 990 total)
You must be logged in to reply to this topic. Login to reply