Add column to table and then update it inside transaction

Question

I am creating a script that will be run in a MS SQL server. This script will run multiple statements and needs to be transactional, if one of the statement fails the overall execution is stopped and any changes are rolled back.

I am having trouble creating this transactional model when issuing ALTER TABLE statements to add columns to a table and then updating the newly added column. In order to access the newly added column right away, I use a GO command to execute the ALTER TABLE statement, and then call my UPDATE statement. The problem I am facing is that I cannot issue a GO command inside an IF statement. The IF statement is important within my transactional model. This is a sample code of the script I am trying to run. Also notice that issuing a GO command, will discard the @errorCode variable, and will need to be declared down in the code before being used (This is not in the code below).

BEGIN TRANSACTION

DECLARE @errorCode INT
SET @errorCode = @@ERROR

-- **********************************
-- * Settings
-- **********************************
IF @errorCode = 0
BEGIN
 BEGIN TRY
  ALTER TABLE Color ADD [CodeID] [uniqueidentifier] NOT NULL DEFAULT ('{00000000-0000-0000-0000-000000000000}')
  GO
 END TRY
 BEGIN CATCH
  SET @errorCode = @@ERROR
 END CATCH
END

IF @errorCode = 0
BEGIN
 BEGIN TRY
  UPDATE Color
  SET CodeID= 'B6D266DC-B305-4153-A7AB-9109962255FC'
  WHERE [Name] = 'Red'
 END TRY
 BEGIN CATCH
  SET @errorCode = @@ERROR
 END CATCH
END

-- **********************************
-- * Check @errorCode to issue a COMMIT or a ROLLBACK
-- **********************************
IF @errorCode = 0
BEGIN
 COMMIT
 PRINT 'Success'
END
ELSE 
BEGIN
 ROLLBACK
 PRINT 'Failure'
END

So what I would like to know is how to go around this problem, issuing ALTER TABLE statements to add a column and then updating that column, all within a script executing as a transactional unit.

Solution 1

GO is not a T-SQL command. Is a batch delimiter. The client tool (SSM, sqlcmd, osql etc) uses it to effectively cut the file at each GO and send to the server the individual batches. So obviously you cannot use GO inside IF, nor can you expect variables to span scope across batches.

Also, you cannot catch exceptions without checking for the XACT_STATE() to ensure the transaction is not doomed.

Using GUIDs for IDs is always at least suspicious.

Using NOT NULL constraints and providing a default 'guid' like '{00000000-0000-0000-0000-000000000000}' also cannot be correct.

Updated:

• Separate the ALTER and UPDATE into two batches.
• Use sqlcmd extensions to break the script on error. This is supported by SSMS when sqlcmd mode is on, sqlcmd, and is trivial to support it in client libraries too: dbutilsqlcmd.
• use XACT_ABORT to force error to interrupt the batch. This is frequently used in maintenance scripts (schema changes). Stored procedures and application logic scripts in general use TRY-CATCH blocks instead, but with proper care: Exception handling and nested transactions.

example script:

:on error exit

set xact_abort on;
go

begin transaction;
go

if columnproperty(object_id('Code'), 'ColorId', 'AllowsNull') is null
begin
    alter table Code add ColorId uniqueidentifier null;
end
go

update Code 
  set ColorId = '...'
  where ...
go

commit;
go

Only a successful script will reach the COMMIT. Any error will abort the script and rollback.

I used COLUMNPROPERTY to check for column existance, you could use any method you like instead (eg. lookup sys.columns).

Solution 2

Orthogonal to Remus's comments, what you can do is execute the update in an sp_executesql.

ALTER TABLE [Table] ADD [Xyz] NVARCHAR(256);

DECLARE @sql NVARCHAR(2048) = 'UPDATE [Table] SET [Xyz] = ''abcd'';';
EXEC sys.sp_executesql @query = @sql;

We've needed to do this when creating upgrade scripts. Usually we just use GO but it has been necessary to do things conditionally.

Solution 3

I almost agree with Remus but you can do this with SET XACT_ABORT ON and XACT_STATE

Basically

• SET XACT_ABORT ON will abort each batch on error and ROLLBACK
• Each batch is separated by GO
• Execution jumps to the next batch on error
• Use XACT_STATE() will test if the transaction is still valid

Tools like Red Gate SQL Compare use this technique

Something like:

SET XACT_ABORT ON
GO
BEGIN TRANSACTION
GO

IF COLUMNPROPERTY(OBJECT_ID('Color'), 'CodeID', ColumnId) IS NULL
   ALTER TABLE Color ADD CodeID [uniqueidentifier] NULL
GO

IF XACT_STATE() = 1
  UPDATE Color
  SET CodeID= 'B6D266DC-B305-4153-A7AB-9109962255FC'
  WHERE [Name] = 'Red'
GO

IF XACT_STATE() = 1
 COMMIT TRAN
--else would be rolled back

I've also removed the default. No value = NULL for GUID values. It's meant to be unique: don't try and set every row to all zeros because it will end in tears...

Solution 4

Have you tried it without the GO?

Normally you should not mix table changes and data changes in the same script.

Solution 5

Another alternative, if you don't want to split the code into separate batches, is to use EXEC to create a nested scope/batch as here

View more solutions

Comments

[Tim]

Your database would be in an ambiguous or otherwise untenable state if only the DDL gets executed? You couldn't do the DDL first, and then your DML, catching any errors on the DML if the DDL had failed?

[Guillermo Gomez]

Thanks Tim, you are right on target! Wish I could vote your comment as a solution too.

[Guillermo Gomez]

Thanks for such quick response. I have already tried that, but it doesn't work. Issuing the UPDATE statement throws an error "Invalid column name 'ColorID'."

[Guillermo Gomez]

Thanks for your response. I am thinking of splitting the script into two. One for DDL and one for DML.

[Guillermo Gomez]

I ended up splitting my script in two files, a DDL and a DML. That way I don't need GO after ALTER statements.

[ZygD]

Nope, the error is related to batch and compilation. At parse time, CodeID does not exist. Nothing to do with statement terminators

[ZygD]

@ggomez: Use SET XACT_ABORT ON like, say, Red gate tools

[Remus Rusanu]

Interesting, use XACT_STATE() as a means to keep state across batches. Didn't know Red Gate does this. Automated generated code can be bulletproofed with IF XACT_STATE() checks on every batch start, but I'm not sure I'd trust developers to keep this discipline for the entire script lifetime... That's why I prefer :on abort exit, even though it relies on client tools to support it.

[ZygD]

@Remus Rusanu: Red Gate uses SET XACT_ABORT ON only and temp tables to span GO. It's easier with XACT_STATE() for simplicity. You're right about developers discipline though. I note your update is along the same lines as my answer too

[Guillermo Gomez]

Thanks gbn, I am including that!

[caesay]

Using GUIDs for IDs is always at least suspicious. It's the default for the PK column of Asp.Net Identity tables, why is it suspicious?

[Yepeekai]

This is great :) but don't use this with powershell invoke-sqlcmd. It will work fine if no errors, but if something fails, not everything will be rolled back and further batches will be executed !!!

[Codure]

This is exactly what I needed. Thank you.

[Asaad Mamoun]

Thank you very much. Saved my day