Question

I have a table called “Schedule” that serves as a queue mechanism for multiple processes/computers hitting the table on a frequent basis.
The goal of the proc is to pick up rows no more then @count of rows, that qualify to be picked up (their LastCompletedProcessingId & LastStartedProcessingId’s must match) and mark them as having been picked up (change LastStartedProcessingId to NEWID()), so that the next process doesn’t try to pick up already marked rows.

My issue is that once in a rare while when the proc is called at very near times by multiple clients, multiple clients end up with the same rows. How is this possible? How do I avoid it? The table itself is not huge. @timeout is not an issue as these things do not take more then 300seconds to process, and i have a log that prior to the processes picking up multiple records, they didnt run over 300seconds. This is running in SQL Azure

Any thoughts of how this can be possible?
Thanks

CREATE PROCEDURE X
    @count int,
    @timeout int = 300
AS
BEGIN
    SET NOCOUNT ON;

    DECLARE @batchId uniqueidentifier
    SELECT @batchId = NEWID()

    BEGIN TRAN

    -- Update rows
    UPDATE Schedule 
    WITH (ROWLOCK)
    SET 
        LastBatchId = @batchId, 
        LastStartedProcessingId = NEWID(), 
        LastStartedProcessingTime = GETDATE()
    WHERE 
        AccountId IN (
            SELECT TOP (@count) AccountId 
            FROM Schedule 
            WHERE 
                (LastStartedProcessingId = LastCompletedProcessingId OR LastCompletedProcessingId IS NULL OR DATEDIFF(SECOND, LastStartedProcessingTime, GETDATE()) > @timeout) AND 
                (LastStartedProcessingTime IS NULL OR DATEDIFF(SECOND, LastStartedProcessingTime, GETDATE()) > Frequency)
            ORDER BY (DATEDIFF(SECOND, LastStartedProcessingTime, GETDATE()) - Frequency) DESC
        )

    -- Return the changed rows
    SELECT AccountId, LastStartedProcessingId, Frequency, LastProcessTime, LastConfigChangeTime
    FROM Schedule 
    WHERE LastBatchId = @batchId

    COMMIT TRAN
END

Answer 1

Instead of using the ROWLOCK hint, run your update with the isolation_level set to SERIALIZABLE.

Also, if you use an OUTPUT clause in your update, you can have the list of affected rows as soon as the update is complete. That means you can end your transaction 1 DML statement sooner and maintain your ACIDity.