更新 (20200226)
有几个 CLR/regex 解决方案可能比我发布的 ngram8k 解决方案更快。我已经听了六年了，但每次测试工具都无一例外地讲述了一个不同的故事。我已经在较早的 cmets 说明中发布了让 CLR Regex 的 Microsoft© MDQ 系列在几分钟内运行的说明。它们由 Microsoft 开发、测试和调整，并附带主数据服务/数据质量服务。我已经用了很多年了，它们很好。
RegexReplace/RegexSplit 与 PatExtract8k/DigitsOnlyEE：1,000,000 行
显然，您不希望 WHEREclause 中包含函数，但是，由于我的正则表达式是生锈的 AF，所以我需要。为了公平竞争，我在 N-Gram 解决方案的 WHERE 子句中对 DigitsOnlyEE 做了同样的事情。
SET NOCOUNT ON;
DBCC FREEPROCCACHE WITH NO_INFOMSGS;
DBCC DROPCLEANBUFFERS WITH NO_INFOMSGS;
SET STATISTICS TIME ON;
DECLARE
@newData BIT = 0,
@string VARCHAR(8000) = '1) Call back from +79005346546, Conversation started<br>Phone: +79005346546<br>Called twice Came from google.com<br>IP: 77.106.46.202 the web page address is xxx.com utm_medium: cpc<br>utm_campaign: 32587871<br>utm_content: 5283041 79005346546 ',
@pattern VARCHAR(50) = '[^0-9()+.-]',
@srchLen INT = 11;
IF @newData = 1
BEGIN
IF OBJECT_ID('tempdb..#strings','U') IS NOT NULL DROP TABLE #strings;
SELECT
StringId = IDENTITY(INT,1,1),
String = REPLICATE(@string,ABS(CHECKSUM(NEWID())%3)+1)
INTO #strings
FROM dbo.rangeAB(1,1000000,1,1) AS r;
END
PRINT CHAR(10)+'Regex/CLR version Serial'+CHAR(10)+REPLICATE('-',90);
SELECT regex.NewString
FROM #strings AS s
CROSS APPLY
(
SELECT STRING_AGG(clr.RegexReplace(f.Token,'[^0-9]','',0),' ')
FROM clr.RegexSplit(s.string,@pattern,N'[0-9()+.-]',0) AS f
WHERE f.IsValid = 1
AND LEN(clr.RegexReplace(f.Token,'[^0-9]','',0)) = @srchLen
) AS regex(NewString);
PRINT CHAR(10)+'NGrams version Serial'+CHAR(10)+REPLICATE('-',90);
SELECT ngramsStuff.NewString
FROM #strings AS s
CROSS APPLY
(
SELECT STRING_AGG(ee.digitsOnly,' ')
FROM samd.patExtract8K(@string,@pattern) AS pe
CROSS APPLY samd.digitsOnlyEE(pe.item) AS ee
WHERE LEN(ee.digitsOnly) = @srchLen
) AS ngramsStuff(NewString)
OPTION (MAXDOP 1);
SET STATISTICS TIME OFF;
GO
测试结果
Regex/CLR version Serial
------------------------------------------------------------------------------------------
SQL Server Execution Times: CPU time = 19918 ms, elapsed time = 12355 ms.
NGrams version Serial
------------------------------------------------------------------------------------------
SQL Server Execution Times: CPU time = 844 ms, elapsed time = 971 ms.
NGrams8k 非常快，不需要您编译新程序集、学习新的编程语言、启用 CLR 功能等......垃圾收集没有问题。即使是 MDS/DQS 附带的 CLR N-GRAMs 功能也无法触及 NGrams8k 的性能（请参阅我文章下的 cmets）。
更新结束
首先获取ngrams8k 的副本并使用它来构建PatExtract8k（本文底部的DDL。）接下来快速热身：
DECLARE
@string VARCHAR(8000) = 'Call me later at 222-3333 or tomorrow at 312.555.2222,
(313)555-6789, or at 1+800-555-4444 before noon. Thanks!',
@pattern VARCHAR(50) = '%[^0-9()+.-]%';
SELECT pe.itemNumber, pe.itemIndex, pe.itemLength, pe.item
FROM samd.patExtract8K(@string,@pattern) AS pe
WHERE pe.itemLength > 1;
返回：
ItemNumber ItemIndex ItemLength Item
----------- ----------- ----------- ----------------
1 18 8 222-3333
2 42 12 312.555.2222
3 91 13 (313)555-6789
4 112 14 1+800-555-4444
请注意，该函数返回匹配的模式、字符串中的位置、项目长度和项目。可以利用前三个属性进行进一步处理，从而将我们带到您的帖子中。注意我的 cmets：
-- First for some easily consumable sample data.
DECLARE @things TABLE (StringId INT IDENTITY, String VARCHAR(8000));
INSERT @things (String)
VALUES
('Call back from +79005346546, Conversation started<br>Phone: +79005346546<br>Called twice Came from google.com<br>IP: 77.106.46.202 the web page address is xxx.com utm_medium: cpc<br>utm_campaign: 32587871<br>utm_content: 5283041 79005346546 '),
('John Smith'),
('xxx@yyy.com'),
('John Smith 8 999 888 77 77');
DECLARE @SrchLen INT = 11;
SELECT
StringId = t.StringId,
ItemIndex = pe.itemIndex,
ItemLength = @SrchLen,
Item = i2.Item
FROM @things AS t
CROSS APPLY samd.patExtract8K(t.String,'[^0-9 ]') AS pe
CROSS APPLY (VALUES(PATINDEX('%'+REPLICATE('[0-9]',@SrchLen), pe.item))) AS i(Idx)
CROSS APPLY (VALUES(SUBSTRING(pe.Item,NULLIF(i.Idx,0),11))) AS ns(NewString)
CROSS APPLY (VALUES(ISNULL(ns.NewString, REPLACE(pe.item,' ','')))) AS i2(Item)
WHERE pe.itemLength >= @SrchLen;
返回：
StringId ItemIndex ItemLength Item
----------- -------------------- ----------- -----------
1 17 11 79005346546
1 62 11 79005346546
1 221 11 79005346546
4 11 11 89998887777
接下来我们可以像这样处理外部行，像这样处理行到列的连接：
WITH t AS
(
SELECT i2.Item, t.StringId
FROM @things AS t
CROSS APPLY samd.patExtract8K(t.String,'[^0-9 ]') AS pe
CROSS APPLY (VALUES(PATINDEX('%'+REPLICATE('[0-9]',@SrchLen), pe.item))) AS i(Idx)
CROSS APPLY (VALUES(SUBSTRING(pe.Item,NULLIF(i.Idx,0),11))) AS ns(NewString)
CROSS APPLY (VALUES(ISNULL(ns.NewString, REPLACE(pe.item,' ','')))) AS i2(Item)
WHERE pe.itemLength >= @SrchLen
)
SELECT
StringId = t2.StringId,
NewString = ISNULL((
SELECT t.item+' '
FROM t
WHERE t.StringId = t2.StringId
FOR XML PATH('')),'')
FROM @things AS t2
LEFT JOIN t AS t1 ON t2.StringId = t1.StringId
GROUP BY t2.StringId;
返回：
StringId NewString
--------- --------------------------------------
1 79005346546 79005346546 79005346546
2
3
4 89998887777
我希望我有更多的时间来了解更多细节，但这比计划的时间要长一些。欢迎提出任何问题。
Patextract：
CREATE FUNCTION samd.patExtract8K
(
@string VARCHAR(8000),
@pattern VARCHAR(50)
)
/*****************************************************************************************
[Description]:
This can be considered a T-SQL inline table valued function (iTVF) equivalent of
Microsoft's mdq.RegexExtract except that:
1. It includes each matching substring's position in the string
2. It accepts varchar(8000) instead of nvarchar(4000) for the input string, varchar(50)
instead of nvarchar(4000) for the pattern
3. The mask parameter is not required and therefore does not exist.
4. You have specify what text we're searching for as an exclusion; e.g. for numeric
characters you should search for '[^0-9]' instead of '[0-9]'.
5. There is is no parameter for naming a "capture group". Using the variable below, both
the following queries will return the same result:
DECLARE @string nvarchar(4000) = N'123 Main Street';
SELECT item FROM samd.patExtract8K(@string, '[^0-9]');
SELECT clr.RegexExtract(@string, N'(?<number>(\d+))(?<street>(.*))', N'number', 1);
Alternatively, you can think of patExtract8K as Chris Morris' PatternSplitCM (found here:
http://www.sqlservercentral.com/articles/String+Manipulation/94365/) but only returns the
rows where [matched]=0. The key benefit of is that it performs substantially better
because you are only returning the number of rows required instead of returning twice as
many rows then filtering out half of them. Furthermore, because we're
The following two sets of queries return the same result:
DECLARE @string varchar(100) = 'xx123xx555xx999';
BEGIN
-- QUERY #1
-- patExtract8K
SELECT ps.itemNumber, ps.item
FROM samd.patExtract8K(@string, '[^0-9]') ps;
-- patternSplitCM
SELECT itemNumber = row_number() over (order by ps.itemNumber), ps.item
FROM dbo.patternSplitCM(@string, '[^0-9]') ps
WHERE [matched] = 0;
-- QUERY #2
SELECT ps.itemNumber, ps.item
FROM samd.patExtract8K(@string, '[0-9]') ps;
SELECT itemNumber = row_number() over (order by itemNumber), item
FROM dbo.patternSplitCM(@string, '[0-9]')
WHERE [matched] = 0;
END;
[Compatibility]:
SQL Server 2008+
[Syntax]:
--===== Autonomous
SELECT pe.ItemNumber, pe.ItemIndex, pe.ItemLength, pe.Item
FROM samd.patExtract8K(@string,@pattern) pe;
--===== Against a table using APPLY
SELECT t.someString, pe.ItemIndex, pe.ItemLength, pe.Item
FROM samd.SomeTable t
CROSS APPLY samd.patExtract8K(t.someString, @pattern) pe;
[Parameters]:
@string = varchar(8000); the input string
@searchString = varchar(50); pattern to search for
[Returns]:
itemNumber = bigint; the instance or ordinal position of the matched substring
itemIndex = bigint; the location of the matched substring inside the input string
itemLength = int; the length of the matched substring
item = varchar(8000); the returned text
[Developer Notes]:
1. Requires NGrams8k
2. patExtract8K does not return any rows on NULL or empty strings. Consider using
OUTER APPLY or append the function with the code below to force the function to return
a row on emply or NULL inputs:
UNION ALL SELECT 1, 0, NULL, @string WHERE nullif(@string,'') IS NULL;
3. patExtract8K is not case sensitive; use a case sensitive collation for
case-sensitive comparisons
4. patExtract8K is deterministic. For more about deterministic functions see:
https://msdn.microsoft.com/en-us/library/ms178091.aspx
5. patExtract8K performs substantially better with a parallel execution plan, often
2-3 times faster. For queries that leverage patextract8K that are not getting a
parallel exeution plan you should consider performance testing using Traceflag 8649
in Development environments and Adam Machanic's make_parallel in production.
[Examples]:
--===== (1) Basic extact all groups of numbers:
WITH temp(id, txt) as
(
SELECT * FROM (values
(1, 'hello 123 fff 1234567 and today;""o999999999 tester 44444444444444 done'),
(2, 'syat 123 ff tyui( 1234567 and today 999999999 tester 777777 done'),
(3, '&**OOOOO=+ + + // ==?76543// and today !!222222\\\tester{}))22222444 done'))t(x,xx)
)
SELECT
[temp.id] = t.id,
pe.itemNumber,
pe.itemIndex,
pe.itemLength,
pe.item
FROM temp AS t
CROSS APPLY samd.patExtract8K(t.txt, '[^0-9]') AS pe;
-----------------------------------------------------------------------------------------
Revision History:
Rev 00 - 20170801 - Initial Development - Alan Burstein
Rev 01 - 20180619 - Complete re-write - Alan Burstein
*****************************************************************************************/
RETURNS TABLE WITH SCHEMABINDING AS RETURN
SELECT itemNumber = ROW_NUMBER() OVER (ORDER BY f.position),
itemIndex = f.position,
itemLength = itemLen.l,
item = SUBSTRING(f.token, 1, itemLen.l)
FROM
(
SELECT ng.position, SUBSTRING(@string,ng.position,DATALENGTH(@string))
FROM samd.NGrams8k(@string, 1) AS ng
WHERE PATINDEX(@pattern, ng.token) < --<< this token does NOT match the pattern
ABS(SIGN(ng.position-1)-1) + --<< are you the first row? OR
PATINDEX(@pattern,SUBSTRING(@string,ng.position-1,1)) --<< always 0 for 1st row
) AS f(position, token)
CROSS APPLY (VALUES(ISNULL(NULLIF(PATINDEX('%'+@pattern+'%',f.token),0),
DATALENGTH(@string)+2-f.position)-1)) AS itemLen(l);
GO