Update, if you are running SQL Server 2012 see: https://stackoverflow.com/a/10309947
The problem is that the SQL Server implementation of the Over clause is somewhat limited.
Oracle (and ANSI-SQL) allow you to do things like:
SELECT somedate, somevalue,
SUM(somevalue) OVER(ORDER BY somedate
ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW)
AS RunningTotal
FROM Table
SQL Server gives you no clean solution to this problem. My gut is telling me that this is one of those rare cases where a cursor is the fastest, though I will have to do some benchmarking on big results.
The update trick is handy but I feel its fairly fragile. It seems that if you are updating a full table then it will proceed in the order of the primary key. So if you set your date as a primary key ascending you will probably be safe. But you are relying on an undocumented SQL Server implementation detail (also if the query ends up being performed by two procs I wonder what will happen, see: MAXDOP):
Full working sample:
drop table #t
create table #t ( ord int primary key, total int, running_total int)
insert #t(ord,total) values (2,20)
-- notice the malicious re-ordering
insert #t(ord,total) values (1,10)
insert #t(ord,total) values (3,10)
insert #t(ord,total) values (4,1)
declare @total int
set @total = 0
update #t set running_total = @total, @total = @total + total
select * from #t
order by ord
ord total running_total
----------- ----------- -------------
1 10 10
2 20 30
3 10 40
4 1 41
You asked for a benchmark this is the lowdown.
The fastest SAFE way of doing this would be the Cursor, it is an order of magnitude faster than the correlated sub-query of cross-join.
The absolute fastest way is the UPDATE trick. My only concern with it is that I am not certain that under all circumstances the update will proceed in a linear way. There is nothing in the query that explicitly says so.
Bottom line, for production code I would go with the cursor.
Test data:
create table #t ( ord int primary key, total int, running_total int)
set nocount on
declare @i int
set @i = 0
begin tran
while @i < 10000
begin
insert #t (ord, total) values (@i, rand() * 100)
set @i = @i +1
end
commit
Test 1:
SELECT ord,total,
(SELECT SUM(total)
FROM #t b
WHERE b.ord <= a.ord) AS b
FROM #t a
-- CPU 11731, Reads 154934, Duration 11135
Test 2:
SELECT a.ord, a.total, SUM(b.total) AS RunningTotal
FROM #t a CROSS JOIN #t b
WHERE (b.ord <= a.ord)
GROUP BY a.ord,a.total
ORDER BY a.ord
-- CPU 16053, Reads 154935, Duration 4647
Test 3:
DECLARE @TotalTable table(ord int primary key, total int, running_total int)
DECLARE forward_cursor CURSOR FAST_FORWARD
FOR
SELECT ord, total
FROM #t
ORDER BY ord
OPEN forward_cursor
DECLARE @running_total int,
@ord int,
@total int
SET @running_total = 0
FETCH NEXT FROM forward_cursor INTO @ord, @total
WHILE (@@FETCH_STATUS = 0)
BEGIN
SET @running_total = @running_total + @total
INSERT @TotalTable VALUES(@ord, @total, @running_total)
FETCH NEXT FROM forward_cursor INTO @ord, @total
END
CLOSE forward_cursor
DEALLOCATE forward_cursor
SELECT * FROM @TotalTable
-- CPU 359, Reads 30392, Duration 496
Test 4:
declare @total int
set @total = 0
update #t set running_total = @total, @total = @total + total
select * from #t
-- CPU 0, Reads 58, Duration 139
Best Answer
"Lock Escalation" is how SQL handles locking for large updates. When SQL is going to change a lot of rows, it's more efficient for the database engine to take fewer, larger locks (e.g. entire table) instead of locking many smaller things (e.g. row locks).
But this can be problematic when you have a huge table, because taking a lock on the entire table may lock out other queries for a long time. That's the tradeoff: many small-granularity locks are slower than fewer (or one) coarse-grained locks, and having multiple queries locking different parts of a table creates the possibility for deadlock if one process is waiting on another.
There is a table-level option,
LOCK_ESCALATION, new in SQL 2008, which allows control of lock escalation. The default, "TABLE" allows locks to escalate all the way to the table level. DISABLE prevents lock escalation to the entire table in most cases. AUTO allows table locks except if the table is partitioned, in which case locks are only made up to the partition level. See this blog post for more info.I suspect that the IDE adds this setting when re-creating a table because TABLE is the default in SQL 2008. Note that LOCK_ESCALATION isn't supported in SQL 2005, so you'll need to strip it if trying to run the script on a 2005 instance. Also, since TABLE is the default, you can safely remove that line when re-running your script.
Also note that, in SQL 2005 before this setting was present, all locks could escalate to table level-- in other words, "TABLE" was the only setting on SQL 2005.