FIX: Update With Self Join May Update Incorrect Number Of Rows (285870)

SYMPTOMS

An UPDATE operation may not modify the correct number of rows if the following conditions are met:

Your query updates a column, ColA, from a table TabA.
The query references TabA again, using an alias TabB.
There is a reference to TabB.ColA somewhere in the query.
None of the columns being updated in TabA are indexed columns.
A hash or merge join is used.

RESOLUTION

SQL Server 2000

To resolve this problem, obtain the latest service pack for Microsoft SQL Server 2000. For additional information, click the following article number to view the article in the Microsoft Knowledge Base:

290211 INF: How to Obtain the Latest SQL Server 2000 Service Pack

SQL Server 7.0

To resolve this problem, obtain the latest service pack for Microsoft SQL Server 7.0. For additional information, click the following article number to view the article in the Microsoft Knowledge Base:

301511 INF: How to Obtain the Latest SQL Server 7.0 Service Pack

NOTE: The following hotfix was created prior to Microsoft SQL Server 7.0 Service Pack 4.

The English version of this fix should have the following file attributes or later:

   Version      File name       Platform
   -------------------------------------

   7.00.978     s70978i.exe     x86
   7.00.978     s70978a.exe     Alpha

NOTE: Because of file dependencies, the most recent hotfix or feature that contains the preceding files may also contain additional files.

MORE INFORMATION

Halloween protection is needed to prevent a situation where the physical location of a row within a table changes due to an UPDATE operation. As a result, the same row may be revisited multiple times within the context of a single logical operation, which should not occur. If the conditions for this particular bug are met, the SQL Server optimizer may not build a plan with appropriate Halloween protection.

The following example demonstrates the problem:

set nocount on
create table test(id int, pid int, fn varchar(256), rn varchar(8))
go
declare @c int
set @c = 1
insert into test values(0, NULL, 'root', 'root')
while @c < 10
begin
   insert into test values(@c, @c-1, NULL, cast(@c as varchar(8)))
   set @c = @c + 1
end
create unique clustered index idx_c_id on test(id)
go
update test
set fn = parent.fn + '/' + test.rn
from test(index=0) , test parent(index=0)
where test.pid = parent.id and test.fn is NULL --  and parent.fn <> ''
option(hash join,force order)

Here are the correct contents of the table after the update:

-- CORRECT results:

id          pid         rn       fn                                                                                                                                                                                                                                                               
----------- ----------- -------- ---- 
0           NULL        root     root
1           0           1        root/1

Here are the results you obtain prior to applying the fix:

-- INCORRECT results:

id          pid         rn       fn                                                                                                                                                                                                                                                               
----------- ----------- -------- ----
0           NULL        root     root
1           0           1        root/1
2           1           2        root/1/2
3           2           3        root/1/2/3
4           3           4        root/1/2/3/4
5           4           5        root/1/2/3/4/5
6           5           6        root/1/2/3/4/5/6
7           6           7        root/1/2/3/4/5/6/7
8           7           8        root/1/2/3/4/5/6/7/8
9           8           9        root/1/2/3/4/5/6/7/8/9