SQL INSERT but avoid duplicates

Question

I want to do some quick inserts but avoid duplicates into a Table. For argument's sake lets call it MarketPrices, I've been experimenting with two ways of doing it but not sure how to benchmark which will be faster.

INSERT INTO MarketPrices (SecurityCode, BuyPrice, SellPrice, IsMarketOpen)
SELECT @SecurityCode, @BuyPrice,  @SellPrice, @IsMarketOpen
EXCEPT
SELECT SecurityCode, BuyPrice, SellPrice, j.bool as IsActive FROM MarketPrices
CROSS JOIN (SELECT 0 as bool UNION SELECT 1 as bool ) as j

OR

DECLARE @MktId int
SET @MktId = (SELECT SecurityId FROM MarketPrices 
              where SecurityCode = @SecurityCode 
              and BuyPrice=@BuyPrice 
              and SellPrice = @SellPrice)

IF (@MktId is NULL)  
BEGIN
    INSERT INTO MarketPrices (SecurityCode, BuyPrice, SellPrice, IsMarketOpen)
    VALUES
    (@SecurityCode,@BuyPrice, @SellPrice, @IsMarketOpen)
END

Assume that @whatever is an input parameter in the stored procedure.

I want to be able to insert a new record for every SecurityCode when the BuyPrice or SellPrice or both are different from every other previous occurance. I don't care about IsMarketOpen.

Is there anything glaringly stupid about either of the above approaches? Is one faster than the other?

Answer 1

EDIT: to prevent race conditions in a concurrent environment, use WITH (UPDLOCK) in the correlated subquery.

---

I think this would be the standard method:

INSERT INTO MarketPrices (SecurityCode, BuyPrice, SellPrice, IsMarketOpen)
SELECT @SecurityCode, @BuyPrice,  @SellPrice, @IsMarketOpen
WHERE NOT EXISTS (
  SELECT * FROM MarketPrices WITH (UPDLOCK)
  WHERE SecurityCode = @SecurityCode 
    AND BuyPrice = @BuyPrice 
    AND SellPrice = @SellPrice
  )

If any of your fields are nullable, you would have to add that to the condition.

Your first method is interesting, but the requirements for EXCEPT have you jumping through hoops. This method is essentially the same, but it gets you around the column matching issue.

Alternatively:

INSERT INTO MarketPrices (SecurityCode, BuyPrice, SellPrice, IsMarketOpen)
SELECT SecurityCode, BuyPrice, SellPrice, @IsMarketOpen
FROM (
  SELECT @SecurityCode, @BuyPrice,  @SellPrice
  EXCEPT
  SELECT SecurityCode, BuyPrice, SellPrice FROM MarketPrices WITH (UPDLOCK)
  ) a (SecurityCode, BuyPrice, SellPrice)

The nice thing about EXCEPT in this instance is that it handles NULLs without any extra coding on your part. To achieve the same thing in first example, you would need to test each pair for NULLs as well as equality, long-hand.

Your second method is ok, but you don't need the variable. See Tomalak's solution, he cleaned it up nicely. Also, you would need to explicitly handle the possibility of concurrent inserts, if that were a concern.

Answer 2

I would go for a semantic solution anytime. Your two proposals seem quite obscure to me (though the latter is better than the former).

IF NOT EXISTS (
  SELECT 1
  FROM   MarketPrices 
  WHERE  SecurityCode  = @SecurityCode 
         AND BuyPrice  = @BuyPrice 
         AND SellPrice = @SellPrice
)  
BEGIN
  INSERT MarketPrices 
    (SecurityCode,   BuyPrice,  SellPrice,  IsMarketOpen)
  VALUES 
    (@SecurityCode, @BuyPrice, @SellPrice, @IsMarketOpen)
END

With a conglomerate index over SecurityCode, BuyPrice, SellPrice the EXISTS query should go reasonably fast.

Benchmarking it is a matter of timing a WHILE loop, I would say. Test it and see for yourself.

Answer 3

Another option: create a unique index on the fields (SecurityCode, BuyPrice, SellPrice) in question, issue a simple insert, and let the database decide whether the records are duplicates. The insert will fail on an attempt to insert a duplicate.

Using code (whether external language or SQL proc) to guarantee uniqueness is not strict enough and will ultimately lead to the very duplicates you hope to prevent.

Answer 4

EDIT: to prevent race conditions in concurrent environments, use WITH (UPDLOCK) in the correlated subquery or EXCEPT'd SELECT. The test script I wrote below doesn't require it, since it uses temporary tables that are only visible to the current connection, but in a real environment, operating against user tables, it would be necessary.

MERGE doesn't require UPDLOCK.

---

Inspired by mcl's answer re: unique index & let the database throw an error, I decided to benchmark conditional inserts vs. try/catch.

The results appear to support the conditional insert over try/catch, but YMMV. It's a very simple scenario (one column, small table, etc), executed on one machine, etc.

Here are the results (SQL Server 2008, build 10.0.1600.2):

duplicates (short table)    
  try/catch:                14440 milliseconds / 100000 inserts
  conditional insert:        2983 milliseconds / 100000 inserts
  except:                    2966 milliseconds / 100000 inserts
  merge:                     2983 milliseconds / 100000 inserts

uniques
  try/catch:                 3920 milliseconds / 100000 inserts
  conditional insert:        3860 milliseconds / 100000 inserts
  except:                    3873 milliseconds / 100000 inserts
  merge:                     3890 milliseconds / 100000 inserts

  straight insert:           3173 milliseconds / 100000 inserts

duplicates (tall table)
  try/catch:                14436 milliseconds / 100000 inserts
  conditional insert:        3063 milliseconds / 100000 inserts
  except:                    3063 milliseconds / 100000 inserts
  merge:                     3030 milliseconds / 100000 inserts

Notice, even on unique inserts, there's slightly more overhead to try/catch than a conditional insert. I wonder if this varies by version, CPU, number of cores, etc.

I did not benchmark the IF conditional inserts, just WHERE. I assume the IF variety would show more overhead, since a) would you have two statements, and b) you would need to wrap the two statements in a transaction and set the isolation level to serializable (!). If someone wanted to test this, you would need to change the temp table to a regular user table (serializable doesn't apply to local temp tables).

Here is the script:

-- tested on SQL 2008.
-- to run on SQL 2005, comment out the statements using MERGE
set nocount on

if object_id('tempdb..#temp') is not null drop table #temp
create table #temp (col1 int primary key)
go

-------------------------------------------------------

-- duplicate insert test against a table w/ 1 record

-------------------------------------------------------

insert #temp values (1)
go

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  begin try 
    insert #temp select @x
  end try
  begin catch end catch
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (short table), try/catch: %i milliseconds / %i inserts',-1,-1,@duration,@y) with nowait
go

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  insert #temp select @x where not exists (select * from #temp where col1 = @x)
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (short table), conditional insert: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  insert #temp select @x except select col1 from #temp
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (short table), except: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go

-- comment this batch out for SQL 2005
declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  merge #temp t using (select @x) s (col1) on t.col1 = s.col1 when not matched by target then insert values (col1);
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (short table), merge: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go

-------------------------------------------------------

-- unique insert test against an initially empty table

-------------------------------------------------------

truncate table #temp
declare @x int, @now datetime, @duration int
select @x = 0, @now = getdate()
while @x < 100000 begin
  set @x = @x+1
  insert #temp select @x
end
set @duration = datediff(ms,@now,getdate())
raiserror('uniques, straight insert: %i milliseconds / %i inserts',-1,-1,@duration, @x) with nowait
go

truncate table #temp
declare @x int, @now datetime, @duration int
select @x = 0, @now = getdate()
while @x < 100000 begin
  set @x = @x+1
  begin try 
    insert #temp select @x
  end try
  begin catch end catch
end
set @duration = datediff(ms,@now,getdate())
raiserror('uniques, try/catch: %i milliseconds / %i inserts',-1,-1,@duration, @x) with nowait
go

truncate table #temp
declare @x int, @now datetime, @duration int
select @x = 0, @now = getdate()
while @x < 100000 begin
  set @x = @x+1
  insert #temp select @x where not exists (select * from #temp where col1 = @x)
end
set @duration = datediff(ms,@now,getdate())
raiserror('uniques, conditional insert: %i milliseconds / %i inserts',-1,-1,@duration, @x) with nowait
go

truncate table #temp
declare @x int, @now datetime, @duration int
select @x = 0, @now = getdate()
while @x < 100000 begin
  set @x = @x+1
  insert #temp select @x except select col1 from #temp
end
set @duration = datediff(ms,@now,getdate())
raiserror('uniques, except: %i milliseconds / %i inserts',-1,-1,@duration, @x) with nowait
go

-- comment this batch out for SQL 2005
truncate table #temp
declare @x int, @now datetime, @duration int
select @x = 1, @now = getdate()
while @x < 100000 begin
  set @x = @x+1
  merge #temp t using (select @x) s (col1) on t.col1 = s.col1 when not matched by target then insert values (col1);
end
set @duration = datediff(ms,@now,getdate())
raiserror('uniques, merge: %i milliseconds / %i inserts',-1,-1,@duration, @x) with nowait
go

-------------------------------------------------------

-- duplicate insert test against a table w/ 100000 records

-------------------------------------------------------

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  begin try 
    insert #temp select @x
  end try
  begin catch end catch
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (tall table), try/catch: %i milliseconds / %i inserts',-1,-1,@duration,@y) with nowait
go

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  insert #temp select @x where not exists (select * from #temp where col1 = @x)
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (tall table), conditional insert: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go

declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  insert #temp select @x except select col1 from #temp
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (tall table), except: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go

-- comment this batch out for SQL 2005
declare @x int, @y int, @now datetime, @duration int
select @x = 1, @y = 0, @now = getdate()
while @y < 100000 begin
  set @y = @y+1
  merge #temp t using (select @x) s (col1) on t.col1 = s.col1 when not matched by target then insert values (col1);
end
set @duration = datediff(ms,@now,getdate())
raiserror('duplicates (tall table), merge: %i milliseconds / %i inserts',-1,-1,@duration, @y) with nowait
go