5.5 How to cope without COMMIT/ROLLBACK

MySQL doesn't support COMMIT-ROLLBACK. The problem is that handling COMMIT-ROLLBACK efficiently would require a completely different table layout than MySQL uses today. MySQL would also need extra threads that do automatic cleanups on the tables and the disk usage would be much higher. This would make MySQL about 2-4 times slower than it is today. MySQL is much faster than almost all other SQL databases (typically at least 2-3 times faster). One of the reasons for this is the lack of COMMIT-ROLLBACK.

For the moment, we are much more for implementing the SQL server language (something like stored procedures). With this you would very seldom really need COMMIT-ROLLBACK. This would also give much better performance.

Loops that need transactions normally can be coded with the help of LOCK TABLES, and you don't need cursors when you can update records on the fly.

We have transactions and cursors on the TODO but not quite prioritized. If we implement these, it will be as an option to CREATE TABLE. That means that COMMIT-ROLLBACK will work only on those tables, so that a speed penalty will be imposed on those table only.

We at TcX have a greater need for a real fast database than a 100% general database. Whenever we find a way to implement these features without any speed loss, we will probably do it. For the moment, there are many more important things to do. Check the TODO for how we prioritize things at the moment. (Customers with higher levels of support can alter this, so things may be reprioritized.)

The current problem is actually ROLLBACK. Without ROLLBACK, you can do any kind of COMMIT action with LOCK TABLES. To support ROLLBACK, MySQL would have to be changed to store all old records that were updated and revert everything back to the starting point if ROLLBACK was issued. For simple cases, this isn't that hard to do (the current isamlog could be used for this purpose), but it would be much more difficult to implement ROLLBACK for ALTER/DROP/CREATE TABLE.

To avoid using ROLLBACK, you can use the following strategy:

  1. Use LOCK TABLES ... to lock all the tables you want to access.
  2. Test conditions.
  3. Update if everything is okay.
  4. Use UNLOCK TABLES to release your locks.

This is usually a much faster method than using transactions with possible ROLLBACKs, although not always. The only situation this solution doesn't handle is when someone kills the threads in the middle of an update. In this case, all locks will be released but some of the updates may not have been executed.

You can also use functions to update records in a single operation. You can get a very efficient application by using the following techniques:

  • Modify fields relative to their current value
  • Update only those fields that actually have changed

For example, when we are doing updates to some customer information, we update only the customer data that have changed and test only that none of the changed data, or data that depend on the changed data, have changed compared to the original row. The test for changed data is done with the WHERE clause in the UPDATE statement. If the record wasn't updated, we give the client a message: "Some of the data you have changed have been changed by another user". Then we show the old row versus the new row in a window, so the user can decide which version of the customer record he should use.

This gives us something that is similar to ``column locking'' but is actually even better, because we only update some of the columns, using values that are relative to their current values. This means that typical UPDATE statements look something like these:

UPDATE tablename SET pay_back=pay_back+'relative change';

UPDATE customer
    address='new address',
    phone='new phone',
    customer_id=id AND address='old address' AND phone='old phone';

As you can see, this is very efficient and works even if another client has changed the values in the pay_back or money_he_owes_us columns.

In many cases, users have wanted ROLLBACK and/or LOCK TABLES for the purpose of managing unique identifiers for some tables. This can be handled much more efficiently by using an AUTO_INCREMENT column and either the SQL function LAST_INSERT_ID() or the C API function mysql_insert_id(). mysql_insert_id().

At TcX, we have never had any need for row-level locking because we have always been able to code around it. Some cases really need row locking, but they are very few. If you want row-level locking, you can use a flag column in the table and do something like this:

UPDATE tbl_name SET row_flag=1 WHERE id=ID;

MySQL returns 1 for the number of affected rows if the row was found and row_flag wasn't already 1 in the original row.

You can think of it as MySQL changed the above query to:

UPDATE tbl_name SET row_flag=1 WHERE id=ID and row_flag <> 1;