问题
We are running many products search on a huge catalog with partially matched barcodes.
We started with a simple like query
select * from products where barcode like '%2345%'
But that takes way too long since it requires a full table scan. We thought a fulltext search will be able to help us here using contains.
select * from products where contains(barcode, '2345')
But, it seems like contains doesn't support finding words that partially contains a text but, only full a word match or a prefix. (But in this example we're looking for '123456').
回答1:
My answer is: @DenisReznik was right :)
ok, let's take a look.
I have worked with barcodes and big catalogs for many years and I was curious about this question.
So I have made some tests on my own.
I have created a table to store test data:
CREATE TABLE [like_test](
[N] [int] NOT NULL PRIMARY KEY,
[barcode] [varchar](40) NULL
)
I know that there are many types of barcodes, some contains only numbers, other contains also letters, and other can be even much complex.
Let's assume our barcode is a random string.
I have filled it with 10 millions records of random alfanumeric data:
insert into like_test
select (select count(*) from like_test)+n, REPLACE(convert(varchar(40), NEWID()), '-', '') barcode
from FN_NUMBERS(10000000)
FN_NUMBERS() is just a function I use in my DBs (a sort of tally_table) to get records quick.
I got 10 million records like that:
N barcode
1 1C333262C2D74E11B688281636FAF0FB
2 3680E11436FC4CBA826E684C0E96E365
3 7763D29BD09F48C58232C7D33551E6C9
Let's declare a var to search for:
declare @s varchar(20) = 'D34F15' -- random alfanumeric string
Let's take a base try with LIKE to compare results to:
select * from like_test where barcode like '%'+@s+'%'
On my workstation it takes 24.4 secs for a full clustered index scan. Very slow.
SSMS suggests to add an index on barcode column:
CREATE NONCLUSTERED INDEX [ix_barcode] ON [like_test] ([barcode]) INCLUDE ([N])
500Mb of index, I retry the select, this time 24.0 secs for the non clustered index seek.. less than 2% better, almost the same result. Very far from the 75% supposed by SSMS. It seems to me this index really doesn't worth it. Maybe my SSD Samsung 840 is making the difference..
For the moment I let the index active.
Let's try the CHARINDEX solution:
select * from like_test where charindex(@s, barcode) > 0
This time it took 23.5 second to complete, not really so much better than LIKE.
Now let's check the @DenisReznik 's suggestion that using the Binary Collation should speed up things.
select * from like_test
where barcode collate Latin1_General_BIN like '%'+@s+'%' collate Latin1_General_BIN
WOW, it seems to work! Only 4.5 secs this is impressive! 5 times better..
So, what about CHARINDEX and Collation toghether? Let's try it:
select * from like_test
where charindex(@s collate Latin1_General_BIN, barcode collate Latin1_General_BIN)>0
Unbelivable! 2.4 secs, 10 times better..
Ok, so far I have realized that CHARINDEX is better than LIKE, and that Binary Collation is better than normal string collation, so from now on I will go on only with CHARINDEX and Collation.
Now, can we do anything else to get even better results? Maybe we can try reduce our very long strings.. a scan is always a scan..
First try, a logical string cut using SUBSTRING to virtually works on barcodes of 8 chars:
select * from like_test
where charindex(
@s collate Latin1_General_BIN,
SUBSTRING(barcode, 12, 8) collate Latin1_General_BIN
)>0
Fantastic! 1.8 seconds.. I have tried both SUBSTRING(barcode, 1, 8) (head of the string) and SUBSTRING(barcode, 12, 8) (middle of the string) with same results.
Then I have tried to phisically reduce the size of the barcode column, almost no difference than using SUBSTRING()
Finally I have tried to drop the index on barcode column and repeated ALL above tests...
I was very surprised to get almost same results, with very little differences.
Index performs 3-5% better, but at cost of 500Mb of disk space and and maintenance cost if the catalog is updated.
Naturally, for a direct key lookup like where barcode = @s with the index it takes 20-50 millisecs, without index we can't get less than 1.1 secs using Collation syntax where barcode collate Latin1_General_BIN = @s collate Latin1_General_BIN
This was interesting.
I hope this helps
回答2:
I often use charindex and just as often have this very debate.
As it turns out, depending on your structure you may actually have a substantial performance boost.
http://cc.davelozinski.com/sql/like-vs-substring-vs-leftright-vs-charindex
回答3:
The good option here for your case - creating your FTS index. Here is how it could be implemented:
1) Create table Terms:
CREATE TABLE Terms
(
Id int IDENTITY NOT NULL,
Term varchar(21) NOT NULL,
CONSTRAINT PK_TERMS PRIMARY KEY (Term),
CONSTRAINT UK_TERMS_ID UNIQUE (Id)
)
Note: index declaration in the table definition is a feature of 2014. If you have a lower version, just bring it out of CREATE TABLE statement and create separately.
2) Cut barcodes to grams, and save each of them to a table Terms. For example: barcode = '123456', your table should have 6 rows for it: '123456', '23456', '3456', '456', '56', '6'.
3) Create table BarcodeIndex:
CREATE TABLE BarcodesIndex
(
TermId int NOT NULL,
BarcodeId int NOT NULL,
CONSTRAINT PK_BARCODESINDEX PRIMARY KEY (TermId, BarcodeId),
CONSTRAINT FK_BARCODESINDEX_TERMID FOREIGN KEY (TermId) REFERENCES Terms (Id),
CONSTRAINT FK_BARCODESINDEX_BARCODEID FOREIGN KEY (BarcodeId) REFERENCES Barcodes (Id)
)
4) Save a pair (TermId, BarcodeId) for the barcode into the table BarcodeIndex. TermId was generated on the second step or exists in the Terms table. BarcodeId - is an identifier of the barcode, stored in Barcodes (or whatever name you use for it) table. For each of the barcodes, there should be 6 rows in the BarcodeIndex table.
5) Select barcodes by their parts using the following query:
SELECT b.* FROM Terms t
INNER JOIN BarcodesIndex bi
ON t.Id = bi.TermId
INNER JOIN Barcodes b
ON bi.BarcodeId = b.Id
WHERE t.Term LIKE 'SomeBarcodePart%'
This solution force all similar parts of barcodes to be stored nearby, so SQL Server will use Index Range Scan strategy to fetch data from the Terms table. Terms in the Terms table should be unique to make this table as small as possible. This could be done in the application logic: check existence -> insert new if a term doesn't exist. Or by setting option IGNORE_DUP_KEY for clustered index of the Terms table. BarcodesIndex table is used to reference Terms and Barcodes.
Please note that foreign keys and constraints in this solution are the points of consideration. Personally, I prefer to have foreign keys, until they hurt me.
回答4:
After further testing and reading and talking with @DenisReznik I think the best option could be to add virtual columns to barcode table to split barcode.
We only need columns for start positions from 2nd to 4th because for the 1st we will use original barcode column and the last I think it is not useful at all (what kind of partial match is 1 char on 6 when 60% of records will match?):
CREATE TABLE [like_test](
[N] [int] NOT NULL PRIMARY KEY,
[barcode] [varchar](6) NOT NULL,
[BC2] AS (substring([BARCODE],(2),(5))),
[BC3] AS (substring([BARCODE],(3),(4))),
[BC4] AS (substring([BARCODE],(4),(3))),
[BC5] AS (substring([BARCODE],(5),(2)))
)
and then to add indexes on this virtual columns:
CREATE NONCLUSTERED INDEX [IX_BC2] ON [like_test2] ([BC2]);
CREATE NONCLUSTERED INDEX [IX_BC3] ON [like_test2] ([BC3]);
CREATE NONCLUSTERED INDEX [IX_BC4] ON [like_test2] ([BC4]);
CREATE NONCLUSTERED INDEX [IX_BC5] ON [like_test2] ([BC5]);
CREATE NONCLUSTERED INDEX [IX_BC6] ON [like_test2] ([barcode]);
now we can simply find partial matches with this query
declare @s varchar(40)
declare @l int
set @s = '654'
set @l = LEN(@s)
select N from like_test
where 1=0
OR ((barcode = @s) and (@l=6)) -- to match full code (rem if not needed)
OR ((barcode like @s+'%') and (@l<6)) -- to match strings up to 5 chars from beginning
or ((BC2 like @s+'%') and (@l<6)) -- to match strings up to 5 chars from 2nd position
or ((BC3 like @s+'%') and (@l<5)) -- to match strings up to 4 chars from 3rd position
or ((BC4 like @s+'%') and (@l<4)) -- to match strings up to 3 chars from 4th position
or ((BC5 like @s+'%') and (@l<3)) -- to match strings up to 2 chars from 5th position
this is HELL fast!
- for search strings of 6 chars 15-20 milliseconds (full code)
- for search strings of 5 chars 25 milliseconds (20-80)
- for search strings of 4 chars 50 milliseconds (40-130)
- for search strings of 3 chars 65 milliseconds (50-150)
- for search strings of 2 chars 200 milliseconds (190-260)
There will be no additional space used for table, but each index will take up to 200Mb (for 1 million barcodes)
PAY ATTENTION
Tested on a Microsoft SQL Server Express (64-bit) and Microsoft SQL Server Enterprise (64-bit) the optimizer of the latter is slight better but the main difference is that:
on express edition you have to extract ONLY the primary key when searching your string, if you add other columns in the SELECT, the optimizer will not use indexes anymore but it will go for full clustered index scan so you will need something like
;with
k as (-- extract only primary key
select N from like_test
where 1=0
OR ((barcode = @s) and (@l=6))
OR ((barcode like @s+'%') and (@l<6))
or ((BC2 like @s+'%') and (@l<6))
or ((BC3 like @s+'%') and (@l<5))
or ((BC4 like @s+'%') and (@l<4))
or ((BC5 like @s+'%') and (@l<3))
)
select N
from like_test t
where exists (select 1 from k where k.n = t.n)
on standard (enterprise) edition you HAVE to go for
select * from like_test -- take a look at the star
where 1=0
OR ((barcode = @s) and (@l=6))
OR ((barcode like @s+'%') and (@l<6))
or ((BC2 like @s+'%') and (@l<6))
or ((BC3 like @s+'%') and (@l<5))
or ((BC4 like @s+'%') and (@l<4))
or ((BC5 like @s+'%') and (@l<3))
回答5:
You do not include many constraints, which means you want to search for string in a string -- and if there was a way to optimized an index to search a string in a string, it would be just built in!
Other things that make it hard to give a specific answer:
It's not clear what "huge" and "too long" mean.
It's not clear as to how your application works. Are you searching in batch as you add a 1,000 new products? Are you allowing a user to enter a partial barcode in a search box?
I can make some suggestions that may or may not be helpful in your case.
Speed up some of the queries
I have a database with lots of licence plates; sometimes an officer wants to search by the last 3-characters of the plate. To support this I store the license plate in reverse, then use LIKE ('ZYX%') to match ABCXYZ. When doing the search, they have the option of a 'contains' search (like you have) which is slow, or an option of doing 'Begins/Ends with' which is super because of the index. This would solve your problem some of the time (which may be good enough), especially if this is a common need.
Parallel Queries
An index works because it organizes data, an index cannot help with a string within a string because there is no organization. Speed seems to be your focus of optimization, so you could store/query your data in a way that searches in parallel. Example: if it takes 10-seconds to sequentially search 10-million rows, then having 10-parallel processes (so process searches 1-million) will take you from 10-seconds to 1-second (kind'a-sort'a). Think of it as scaling out. There are various options for this, within your single SQL Instance (try data partitioning) or across multiple SQL Servers (if that's an option).
BONUS: If you're not on a RAID setup, that can help with reads since it's a effectively of reading in parallel.
Reduce a bottleneck
One reason searching "huge" datasets take "too long" is because all that data needs to be read from the disk, which is always slow. You can skip-the-disk, and use InMemory Tables. Since "huge" isn't defined, this may not work.
回答6:
UPDATED:
We know from that FULL-TEXT searches can be used for the following:
Full-Text Search - MSDN
- One or more specific words or phrases (simple term)
- A word or a phrase where the words begin with specified text (prefix term)
- Inflectional forms of a specific word (generation term)
- A word or phrase close to another word or phrase (proximity term)
- Synonymous forms of a specific word (thesaurus)
- Words or phrases using weighted values (weighted term)
Are any of these fulfilled by your query requirements? If you are having to search for patterns as you described, without an consistent pattern (such as '1%'), then there may not be a way for SQL to use a SARG.
- You could use
Booleanstatements
Coming from a C++ perspective, B-Trees are accessed from Pre-Order, In-Order, and Post-Order traversals and utilize Boolean statements to search the B-Tree. Processed much faster than string comparisons, booleans offer at the least an improved performance.
We can see this in the following two options:
PATINDEX
- Only if your column is not numeric, as PATINDEX is designed for strings.
- Returns an integer (like CHARINDEX) which is easier to process than strings.
CHARINDEX is a solution
- CHARINDEX has no problem searching INTs and again, returns a number.
- May require some extra cases built in (i.e. first number is always ignored), but you can add them like so:
CHARINDEX('200', barcode) > 1.
Proof of what I am saying, let us go back to the old [AdventureWorks2012].[Production].[TransactionHistory]. We have TransactionID which contains the number of the items we want, and lets for fun assume you want every transactionID that has 200 at the end.
-- WITH LIKE
SELECT TOP 1000 [TransactionID]
,[ProductID]
,[ReferenceOrderID]
,[ReferenceOrderLineID]
,[TransactionDate]
,[TransactionType]
,[Quantity]
,[ActualCost]
,[ModifiedDate]
FROM [AdventureWorks2012].[Production].[TransactionHistory]
WHERE TransactionID LIKE '%200'
-- WITH CHARINDEX(<delimiter>, <column>) > 3
SELECT TOP 1000 [TransactionID]
,[ProductID]
,[ReferenceOrderID]
,[ReferenceOrderLineID]
,[TransactionDate]
,[TransactionType]
,[Quantity]
,[ActualCost]
,[ModifiedDate]
FROM [AdventureWorks2012].[Production].[TransactionHistory]
WHERE CHARINDEX('200', TransactionID) > 3
Note CHARINDEX removes the value 200200 in the search, so you may need to adjust your code appropriately. But look at the results:
- Clearly, booleans and numbers are faster comparisons.
- LIKE uses string comparisons, which again is much slower to process.
I was a bit surprised at the size of the difference, but the fundamentals are the same. Integers and Boolean statements are always faster to process than string comparisons.
回答7:
Full-text is aimed for bigger texts, let's say texts with more than about 100 chars. You can use LIKE '%string%'. (However it depends how the barcode column is defined.) Do you have an index for barcode? If not, then create one and it will improve your query.
回答8:
First make the index on column on which you have to put as where clause .
Secondly for the datatype of the column which are used in where clause make them as Char in place of Varchar which will save you some space,in the table and in the indexes that will include that column. varchar(1) column needs one more byte over char(1)
Do pull only the number of columns you need try to avoid * , be specific to number of columns you wish to select. Don't write as
select * from products
In place of it write as
Select Col1, Col2 from products with (Nolock)
来源：https://stackoverflow.com/questions/37677204/using-sqlserver-contains-for-partial-words