run-periodic-tasks errors and data loss

At some point during the day on Tuesday, my server crashed - hard. I am not sure run-periodic-tasks was the culprit for the crash, but considering the state of my MT installation afterwards, I wouldn't at all be surprised.

We were seeing decent traffic (15K page views) when the server suddenly stopped responding. On reboot, we found that the system logger had even died, so tracking down the exact nature of the crash would be impossible.

Run-periodic-tasks no longer works and it almost crashed my server again because 1 instance of the job ran for over 6 hours until it had eaten all the available memory on the server and I had to kill the process.

Now, when I run RPT with --verbose, I get messages like this:

prepare_cached(SELECT role_id, role_created_by, role_created_on, role_description,
role_is_system, role_modified_by, role_modified_on, role_name, role_permissions, 
role_role_mask, role_role_mask2, role_role_mask3, role_role_mask4
FROM mt_role, mt_association
WHERE (association_author_id = ?) AND (association_blog_id = ?) AND 
(association_type = ?) AND (role_id = association_role_id)
) statement handle DBI::st=HASH(0xdcda75c) still Active at /var/www/www.bilerico.com
/cgi-bin/mt/extlib/Data/ObjectDriver/Driver/DBI.pm line 117

The job never completes and just keeps running until it saps all the memory from the server. I also get errors in the MT Error Log that say my /tmp directory isn't writable while the RPT job is running. My /tmp is writable and has the sticky bit set just like it should.

Also, this is the second major crash I've had just like this one since upgrading to 4.1x. Both times, I've lost settings and custom templates. For some reason, my MT installation decided to reset the "registration" settings to settings I've never used (turned off MT registrations, turned ON TypeKey) on every blog on the system. I also lost my "Comment Response" template.

This has happened to me twice, like I said, but never has RPT locked up like it is now.

There are no jobs in the ts_jobs table, so these are some kind of system event or scheduled entry event that it is trying to handle.

How do I clear out the queue for RPT if there is nothing in ts_jobs? The only plugin I have that I'm aware of using RPT is VisitorStats, but I tried disabling that plugin and still get the errors.