Determine Old primary key in a SQL Trigger

Question

I've done this before somewhere I'm sure of it!

I have a SQL Server 2000 table that I need to log changes to fields on updates and inserts into a second Logging table. A simplified version of the structure I'm using is below:

MainTable
ID varchar(10) PRIMARY KEY
DESCRIPTION varchar(50)

LogTable
OLDID varchar(10)
NEWID varchar(10)

For any other field something like this would work great:

Select i.DESCRIPTION As New, d.DESCRIPTION As Old 
From Inserted i
LEFT JOIN Deleted d On i.ID=d.ID

...But obviously the join would fail if ID was changed.

I cannot modify the Tables in way, the only power I have in this database is to create a trigger.

Alternatively is there someone who can teach me time travelling and I'll go back into the past and ask myself back then how I did this? Cheers :)

---

Edit:

I think I need to clarify a few things here. This is not actually my database, it is a pre-existing system that I have almost no control of, other than writing this trigger.

My question is how can I retrieve the old primary key if said primary key was changed. I don't need to be told that I shouldn't change the primary key or about chasing up foreign keys etc. That's not my problem :)

Answer 1

Is it possible to assume that the INSERTED and DELETED tables presented to you in a trigger are guaranteed to be in the same order?

Answer 2

DECLARE @OldKey int, @NewKey int;

SELECT @Oldkey = [ID] FROM DELETED;
SELECT @NewKey = [ID] FROM INSERTED;

This only works if you have a single row. Otherwise you have no "anchor" to link old and new rows. So check in your trigger for > 1 in INSERTED.

Answer 3

I don't think it's possible. Imagine if you have 4 rows in the table:

1  Val1
2  Val2
3  Val3
4  Val4

Now issue the following update:

UPDATE MainTable SET
ID = CASE ID WHEN 1 THEN 2 WHEN 2 THEN 1 ELSE ID END
Description = CASE ID WHEN 3 THEN 'Val4' WHEN 4 THEN 'Val3' ELSE Description END

Now, how are you going to distinguish between what happened to rows 1 & 2 and what happened to rows 3 & 4. And more importantly, can you describe what's different between them? All of the stuff that tells you which columns have been updated won't help you.

If it's possible in this case that there's an additional key on the table (e.g. Description is UNIQUE), and your update rules allow it, you could write the trigger to prevent simultaneous updates to both keys, and then you can use whichever key hasn't been updated to correlate the two tables.

Answer 4

If you must handle multiple-row inserts/updates, and there's no alternate key that's guaranteed not to change, the only way I can see to do this is to use an INSTEAD OF trigger. For example, in the trigger you could break the original insert/update command into one command per row, grabbing each old id before you insert/update.

Answer 5

Within triggers in SQL Server you have access to two tables: deleted and inserted. Both of these have already been mentioned. Here's how they function depending on what action the trigger is firing on:

INSERT OPERATION

• deleted - not used
• inserted - contains the new rows being added to the table

DELETE OPERATION

• deleted - contains the rows being removed from the table
• inserted - not used

UPDATE OPERATION

• deleted - contains the rows as they would exist before the UPDATE operation
• inserted - contains the rows as they would exist after the UPDATE operation

These function in every way like tables. Therefore, it is entirely possible to use a row based operation such as something like the following (Operation exists only on the audit table, as does DateChanged):

INSERT INTO MyAuditTable
(ID, FirstColumn, SecondColumn, ThirdColumn, Operation, DateChanged)
VALUES
SELECT ID, FirstColumn, SecondColumn, ThirdColumn, 'Update-Before', GETDATE()
FROM deleted
UNION ALL
SELECT ID, FirstColumn, SecondColumn, ThirdColumn, 'Update-After', GETDATE()
FROM inserted

Answer 6

----new---- add an identity column to the table that the application can not change, you can then use that new column to join the inserted to the deleted tables within the trigger:

ALTER TABLE YourTableName ADD
    PrivateID int NOT NULL IDENTITY (1, 1)
GO

----old---- Don't ever update/change key values. How can you do this and fix all of your foreign keys?

I wouldn't recommend ever using a trigger that can't handle a set of rows.

If you must change the key, insert a new row with the proper new key and values, use SCOPE_IDENTITY() if that is what your are doing. Delete the old row. Log for the old row that it was changed to the new row's key, which you should now have. I hope there is no foreign key on the changed key in your log...

Answer 7

You can create a new identity column on table MainTable (named for example correlationid) and correlate inserted and deleted tables using this column. This new column should be transparent for existing code.

INSERT INTO LOG(OLDID, NEWID)
SELECT deleted.id AS OLDID, inserted.id AS NEWID
FROM inserted 
INNER JOIN deleted 
    ON inserted.correlationid = deleted.correlationid

Pay attention, you could insert duplicate records in the log table.

Answer 8

Of course nobody should be changing the primary key on the table -- but that is exactly what triggers are supposed to be for (in part), is to keep people from doing things they shouldn't do. It's a trivial task in Oracle or MySQL to write a trigger that intercepts changes to primary keys and stops them, but not at all easy in SQL Server.

What you of course would love to be able to do would be to simply do something like this:

if exists
  (
  select *
    from inserted changed
           join deleted old
   where changed.rowID = old.rowID
     and changed.id != old.id
  )
... [roll it all back]

Which is why people go out googling for the SQL Server equivalent of ROWID. Well, SQL Server doesn't have it; so you have to come up with another approach.

A fast, but sadly not bombproof, version is to write an instead of update trigger that looks to see whether any of the inserted rows have a primary key not found in the updated table or vice versa. This would catch MOST, but not all, of the errors:

if exists
  (
  select *
    from inserted lost
           left join updated match
             on match.id = lost.id
   where match.id is null
  union
  select *
    from deleted new
           left join inserted match
             on match.id = new.id
    where match.id is null
  )
  -- roll it all back

But this still doesn't catch an update like...

update myTable
   set id = case
              when id = 1 then 2 
              when id = 2 then 1
              else id
              end

Now, I've tried making the assumption that the inserted and deleted tables are ordered in such a way that cursoring through the inserted and deleted tables simultaneously will give you properly matching rows. And this APPEARS to work. In effect you turn the trigger into the equivalent of the for-each-row triggers available in Oracle and mandatory in MySQL...but I would imagine the performance will be bad on massive updates since this is not native behavior to SQL Server. Also it depends upon an assumption that I can't actually find documented anywhere and so am reluctant to depend on. But code structured that way APPEARS to work properly on my SQL Server 2008 R2 installation. The script at the end of this post highlights both the behavior of the fast-but-not-bombproof solution and the behavior of the second, pseudo-Oracle solution.

If anybody could point me to someplace where my assumption is documented and guaranteed by Microsoft I'd be a very grateful guy...

begin try
  drop table kpTest;
end try
begin catch
end catch
go

create table kpTest( id int primary key, name nvarchar(10) )
go

begin try
  drop trigger kpTest_ioU;
end try
begin catch
end catch
go

create trigger kpTest_ioU on kpTest
instead of update
as
begin
  if exists
    (
    select *
      from inserted lost
             left join deleted match
               on match.id = lost.id
     where match.id is null
    union
    select *
      from deleted new
             left join inserted match
               on match.id = new.id
      where match.id is null
    )
      raisError( 'Changed primary key', 16, 1 )
  else
    update kpTest
       set name = i.name
      from kpTest
             join inserted i
               on i.id = kpTest.id
    ;
end
go

insert into kpTest( id, name ) values( 0, 'zero' );
insert into kpTest( id, name ) values( 1, 'one' );
insert into kpTest( id, name ) values( 2, 'two' );
insert into kpTest( id, name ) values( 3, 'three' );

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/

-- This throws an error, appropriately
update kpTest set id = 5, name = 'FIVE' where id = 1
go

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/

-- This allows the change, inappropriately
update kpTest 
   set id = case   
              when id = 1 then 2
              when id = 2 then 1
              else id
              end
     , name = UPPER( name )
go

select * from kpTest

/*
0   ZERO
1   TWO   -- WRONG WRONG WRONG
2   ONE   -- WRONG WRONG WRONG
3   THREE
*/

-- Put it back
update kpTest 
   set id = case   
              when id = 1 then 2
              when id = 2 then 1
              else id
              end
     , name = LOWER( name )
go

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/

drop trigger kpTest_ioU
go

create trigger kpTest_ioU on kpTest
instead of update
as
begin
  declare newIDs cursor for select id, name from inserted;
  declare oldIDs cursor for select id from deleted;
  declare @thisOldID int;
  declare @thisNewID int;
  declare @thisNewName nvarchar(10);
  declare @errorFound int;
  set @errorFound = 0;
  open newIDs;
  open oldIDs;
  fetch newIDs into @thisNewID, @thisNewName;
  fetch oldIDs into @thisOldID;
  while @@FETCH_STATUS = 0 and @errorFound = 0
    begin
      if @thisNewID != @thisOldID
        begin
          set @errorFound = 1;
          close newIDs;
          deallocate newIDs;
          close oldIDs;
          deallocate oldIDs;
          raisError( 'Primary key changed', 16, 1 );
        end
      else
        begin
          update kpTest
             set name = @thisNewName
           where id = @thisNewID
          ;
          fetch newIDs into @thisNewID, @thisNewName;
          fetch oldIDs into @thisOldID;
        end
    end;
  if @errorFound = 0
    begin
      close newIDs;
      deallocate newIDs;
      close oldIDs;
      deallocate oldIDs;
    end
end
go

-- Succeeds, appropriately
update kpTest
   set name = UPPER( name )
go

select * from kpTest;

/*
0   ZERO
1   ONE
2   TWO
3   THREE
*/

-- Succeeds, appropriately
update kpTest
   set name = LOWER( name )
go

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/


-- Fails, appropriately
update kpTest 
   set id = case   
              when id = 1 then 2
              when id = 2 then 1
              else id
              end
go

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/

-- Fails, appropriately
update kpTest 
   set id = id + 1
go

select * from kpTest;

/*
0   zero
1   one
2   two
3   three
*/

-- Succeeds, appropriately
update kpTest 
   set id = id, name = UPPER( name )
go

select * from kpTest;

/*
0   ZERO
1   ONE
2   TWO
3   THREE
*/

drop table kpTest
go