I've found that if PHP exits due to a code bug during a transaction, an InnoDB table can remain locked until Apache is restarted.
The simple test is to start a transaction by setting $mysqli_obj->autocommit(false) and executing an insert statement. Before getting to a $mysqli_obj->commit statement - have a runtime code bug bomb PHP. You check the database, no insert happened (you assume a rollback occurred) .. and you go fix the bug, and try again... but this time the script takes about 50 seconds to timeout - the insert statement returning with a "1205 - Lock wait timeout exceeded; try restarting transaction". No rollback occurred. And this error will not go away until you restart Apache - for whatever reason, the resources are not released until the process is killed.
I found that an 'exit', instead of a PHP code bug, will not cause a problem. So there is an auto-rollback mechanism in place - it just fails miserably when PHP dies unexpectantly. Having to restarting apache is a pretty drastic measure to overcome a code bug.
To avoid this problem, I use "register_shutdown_function()" when I start a transaction, and set a flag to indicate a transaction is in process (because there is no unregister_shutdown_function()). See below. So the __shutdown_check() routine (I beleive it needs to be public) is called when the script bombs - which is able to invoke the rollback().
these are just the relevant bits to give u an idea...
<?php
public function begin_transaction() {
$ret = $this->mysqli_obj->autocommit(false);
$this->transaction_in_progress = true;
register_shutdown_function(array($this, "__shutdown_check"));
}
public function __shutdown_check() {
if ($this->transaction_in_progress) {
$this->rollback();
}
}
public function commit() {
$ret = $this->mysqli_obj->commit();
$this->transaction_in_progress = false;
}
public function rollback() {
$ret = $this->mysqli_obj->rollback();
$this->transaction_in_progress = false;
}
?>
True for PHP 5.1.6 + MySQL 5.0.24a.