SQL Server "AFTER INSERT" trigger doesn't see the just-inserted row

Consider this trigger:

ALTER TRIGGER myTrigger 
   ON someTable 
   AFTER INSERT
AS BEGIN
  DELETE FROM someTable
         WHERE ISNUMERIC(someField) = 1
END

I've got a table, someTable, and I'm trying to prevent people from inserting bad records. For the purpose of this question, a bad record has a field "someField" that is all numeric.

Of course, the right way to do this is NOT with a trigger, but I don't control the source code... just the SQL database. So I can't really prevent the insertion of the bad row, but I can delete it right away, which is good enough for my needs.

The trigger works, with one problem... when it fires, it never seems to delete the just-inserted bad record... it deletes any OLD bad records, but it doesn't delete the just-inserted bad record. So there's often one bad record floating around that isn't deleted until somebody else comes along and does another INSERT.

Is this a problem in my understanding of triggers? Are newly-inserted rows not yet committed while the trigger is running?

Triggers cannot modify the changed data (Inserted or Deleted) otherwise you could get infinite recursion as the changes invoked the trigger again. One option would be for the trigger to roll back the transaction.

Edit: The reason for this is that the standard for SQL is that inserted and deleted rows cannot be modified by the trigger. The underlying reason for is that the modifications could cause infinite recursion. In the general case, this evaluation could involve multiple triggers in a mutually recursive cascade. Having a system intelligently decide whether to allow such updates is computationally intractable, essentially a variation on the halting problem.

The accepted solution to this is not to permit the trigger to alter the changing data, although it can roll back the transaction.

create table Foo (
       FooID int
      ,SomeField varchar (10)
)
go

create trigger FooInsert
    on Foo after insert as
    begin
        delete inserted
         where isnumeric (SomeField) = 1
    end
go


Msg 286, Level 16, State 1, Procedure FooInsert, Line 5
The logical tables INSERTED and DELETED cannot be updated.

Something like this will roll back the transaction.

create table Foo (
       FooID int
      ,SomeField varchar (10)
)
go

create trigger FooInsert
    on Foo for insert as
    if exists (
       select 1
         from inserted 
        where isnumeric (SomeField) = 1) begin
              rollback transaction
    end
go

insert Foo values (1, '1')

Msg 3609, Level 16, State 1, Line 1
The transaction ended in the trigger. The batch has been aborted.

You can reverse the logic. Instead of deleting an invalid row after it has been inserted, write an INSTEAD OF trigger to insert only if you verify the row is valid.

CREATE TRIGGER mytrigger ON sometable
INSTEAD OF INSERT
AS BEGIN
  DECLARE @isnum TINYINT;

  SELECT @isnum = ISNUMERIC(somefield) FROM inserted;

  IF (@isnum = 1)
    INSERT INTO sometable SELECT * FROM inserted;
  ELSE
    RAISERROR('somefield must be numeric', 16, 1)
      WITH SETERROR;
END

If your application doesn't want to handle errors (as Joel says is the case in his app), then don't RAISERROR. Just make the trigger silently not do an insert that isn't valid.

I ran this on SQL Server Express 2005 and it works. Note that INSTEAD OF triggers do not cause recursion if you insert into the same table for which the trigger is defined.

Here's my modified version of Bill's code:

CREATE TRIGGER mytrigger ON sometable
INSTEAD OF INSERT
AS BEGIN
  INSERT INTO sometable SELECT * FROM inserted WHERE ISNUMERIC(somefield) = 1 FROM inserted;
  INSERT INTO sometableRejects SELECT * FROM inserted WHERE ISNUMERIC(somefield) = 0 FROM inserted;
END

This lets the insert always succeed, and any bogus records get thrown into your sometableRejects where you can handle them later. It's important to make your rejects table use nvarchar fields for everything - not ints, tinyints, etc - because if they're getting rejected, it's because the data isn't what you expected it to be.

This also solves the multiple-record insert problem, which will cause Bill's trigger to fail. If you insert ten records simultaneously (like if you do a select-insert-into) and just one of them is bogus, Bill's trigger would have flagged all of them as bad. This handles any number of good and bad records.

I used this trick on a data warehousing project where the inserting application had no idea whether the business logic was any good, and we did the business logic in triggers instead. Truly nasty for performance, but if you can't let the insert fail, it does work.