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').
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
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
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.
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))
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.
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 Boolean statements
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.
I'm late to the game but here's another way to get a full-text like index in the spirit of #MtwStark's second answer.
This is a solution using a search table join
drop table if exists #numbers
select top 10000 row_number() over(order by t1.number) as n
into #numbers
from master..spt_values t1
cross join master..spt_values t2
drop table if exists [like_test]
create TABLE [like_test](
[N] INT IDENTITY(1,1) not null,
[barcode] [varchar](40) not null,
constraint pk_liketest primary key ([N])
)
insert into dbo.like_test (barcode)
select top (1000000) replace(convert(varchar(40), NEWID()), '-', '') barcode
from #numbers t,#numbers t2
drop table if exists barcodesearch
select distinct ps.n, trim(substring(ps.barcode,ty.n,100)) as searchstring
into barcodesearch
from like_test ps
inner join #numbers ty on ty.n < 40
where len(ps.barcode) > ty.n
create clustered index idx_barcode_search_index on barcodesearch (searchstring)
The final search should look like this:
declare #s varchar(20) = 'D34F15'
select distinct lt.* from dbo.like_test lt
inner join barcodesearch bs on bs.N = lt.N
where bs.searchstring like #s+'%'
If you have the option of full-text searching, you can speed this up even further by adding the full-text search column directly to the barcode table
drop table if exists #liketestupdates
select n, string_agg(searchstring, ' ')
within group (order by reverse(searchstring)) as searchstring
into #liketestupdates
from barcodesearch
group by n
alter table dbo.like_test add search_column varchar(559)
update lt
set search_column = searchstring
from like_test lt
inner join #liketestupdates lu on lu.n = lt.n
CREATE FULLTEXT CATALOG ftcatalog as default;
create fulltext index on dbo.like_test ( search_column )
key index pk_liketest
The final full-text search would look like this:
declare #s varchar(20) = 'D34F15'
set #s = '"*' + #s + '*"'
select n,barcode from dbo.like_test where contains(search_column, #s)
I understand that Estimated Costs aren't the best measure of expected performance but the number's aren't wildly off here.
With the search table join, the Estimated Subtree Cost is 2.13
With the full-text search, the Estimated Subtree Cost is 0.008
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.
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)
Which one of the following queries is faster (LIKE vs CONTAINS)?
SELECT * FROM table WHERE Column LIKE '%test%';
or
SELECT * FROM table WHERE Contains(Column, "test");
The second (assuming you means CONTAINS, and actually put it in a valid query) should be faster, because it can use some form of index (in this case, a full text index). Of course, this form of query is only available if the column is in a full text index. If it isn't, then only the first form is available.
The first query, using LIKE, will be unable to use an index, since it starts with a wildcard, so will always require a full table scan.
The CONTAINS query should be:
SELECT * FROM table WHERE CONTAINS(Column, 'test');
Having run both queries on a SQL Server 2012 instance, I can confirm the first query was fastest in my case.
The query with the LIKE keyword showed a clustered index scan.
The CONTAINS also had a clustered index scan with additional operators for the full text match and a merge join.
I think that CONTAINS took longer and used Merge because you had a dash("-") in your query adventure-works.com.
The dash is a break word so the CONTAINS searched the full-text index for adventure and than it searched for works.com and merged the results.
Also try changing from this:
SELECT * FROM table WHERE Contains(Column, "test") > 0;
To this:
SELECT * FROM table WHERE Contains(Column, '"*test*"') > 0;
The former will find records with values like "this is a test" and "a test-case is the plan".
The latter will also find records with values like "i am testing this" and "this is the greatest".
I didn't understand actually what is going on with "Contains" keyword. I set a full text index on a column. I run some queries on the table.
Like returns 450.518 rows but contains not and like's result is correct
SELECT COL FROM TBL WHERE COL LIKE '%41%' --450.518 rows
SELECT COL FROM TBL WHERE CONTAINS(COL,N'41') ---40 rows
SELECT COL FROM TBL WHERE CONTAINS(COL,N'"*41*"') -- 220.364 rows
I have a SQL Server table and I'm trying to make sense of fulltext searching :-)
I have set up a fulltext catalog and a fulltext index on a table Entry, which contains among other columns a VARCHAR(20) column called VPN-ID.
There are about 200'000 rows in that table, and the VPN-ID column has values such as:
VPN-000-359-90
VPN-000-363-85
VPN-000-362-07
VPN-000-362-91
VPN-000-355-55
VPN-000-368-36
VPN-000-356-90
Now I'm trying to find rows in that table with a fulltext enabled search.
When I do
SELECT (list of columns)
FROM dbo.Entry
WHERE CONTAINS(*, 'VPN-000-362-07')
everything's fine and dandy and my rows are returned.
When I start searching with a wildcard like this:
SELECT (list of columns)
FROM dbo.Entry
WHERE CONTAINS(*, 'VPN-000-362-%')
I am getting results and everything seems fine.
HOWEVER: when I searching like this:
SELECT (list of columns)
FROM dbo.Entry
WHERE CONTAINS(*, 'VPN-000-36%')
suddenly I get no results back at all..... even though there are clearly rows that match that search criteria...
Any ideas why?? What other "surprises" might fulltext search have in store for me? :-)
Update: to create my fulltext catalog I used:
CREATE FULLTEXT CATALOG MyCatalog WITH ACCENT_SENSITIVITY = OFF
and to create the fulltext index on my table, I used
CREATE FULLTEXT INDEX
ON dbo.Entry(list of columns)
KEY INDEX PK_Entry
I tried to avoid any "oddball" options as much a I could.
Update #2: after a bit more investigation, it appears as if SQL Server Fulltext search somehow interprets my dashes inside the strings as separators....
While this query returns nothing:
SELECT (list of columns)
FROM dbo.Entry
WHERE CONTAINS(*, '"VPN-000-362*"')
this one does (splitting up the search term on the dashes):
SELECT (list of columns)
FROM dbo.Entry
WHERE CONTAINS(*, ' "VPN" AND "000" AND "362*"')
OK - seems a bit odd that a dash appears to result in a splitting up that somehow doesn't work.....
which Language for Word Breaker do you use? Have you tried Neutral?
EDIT:
in adition you should use WHERE CONTAINS([Column], '"text*"'). See MSDN for more information on Prefix Searches:
C. Using CONTAINS with
The following example returns all
product names with at least one word
starting with the prefix chain in the
Name column.
USE AdventureWorks2008R2;
GO
SELECT Name
FROM Production.Product
WHERE CONTAINS(Name, ' "Chain*" ');
GO
btw ... similar question here and here
Just wondering, but why don't you just do this:
SELECT (list of columns)
FROM dbo.Entry
WHERE [VPN-ID] LIKE 'VPN-000-36%'
It seems to me that fulltext search is not the right tool for the job. Just use a normal index on that column.
I am doing a MS SQL Server Full Text Search query. I need to escape special characters so I can search on a specific term that contains special characters. Is there a built-in function to escape a full text search string? If not, how would you do it?
Bad news: there's no way. Good news: you don't need it (as it won't help anyway).
I've faced similar issue on one of my projects. My understanding is that while building full-text index, SQL Server treats all special characters as word delimiters and hence:
Your word with such a character is represented as two (or more) words in full-text index.
These character(s) are stripped away and don't appear in an index.
Consider we have the following table with a corresponding full-text index for it (which is skipped):
CREATE TABLE [dbo].[ActicleTable]
(
[Id] int identity(1,1) not null primary key,
[ActicleBody] varchar(max) not null
);
Consider later we add rows to the table:
INSERT INTO [ActicleTable] values ('digitally improvements folders')
INSERT INTO [ActicleTable] values ('digital"ly improve{ments} fold(ers)')
Try searching:
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'digitally')
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'improvements')
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'folders')
and
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'digital')
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'improve')
SELECT * FROM [ArticleTable] WHERE CONTAINS(*, 'fold')
First group of conditions will match first row (and not the second) while the second group will match second row only.
Unfortunately I could not find a link to MSDN (or something) where such behaviour is clearly stated. But I've found an official article that tells how to convert quotation marks for full-text search queries, which is [implicitly] aligned with the above described algorithm.