Detecting MySQL server problems automaticallyThu, Sep 6, 2012 in Databases
I previously blogged about work I was doing on automatically finding problems in a MySQL server, with no hardcoded thresholds or predetermined indicators of what is “bad behavior.” I had to pause my studies on that for a while, due to time constraints. I’ve recently been able to resume and I’m happy to report that I’m making good progress.
One of the things I’ve done is a survey of existing literature on this subject. It turns out that the abnormality-detection techniques I’ve developed over the years are well-known in the operations research field. I reinvented some classic techniques used in Statistical Process Control (SPC). These include Shewhart Control Charts, exponentially weighted moving averages, and Holt-Winters forecasting. However, I was never satisfied with these approaches. They are simultaneously overly simplistic and too sensitive, so they produce false positives and false negatives when applied to MySQL status metrics. I suspect the same thing would happen in most server systems.
Given that the existing techniques I’ve found are inadequate, I’ve developed several more that satisfy me. I am currently testing them on a wide variety of real-world data to fine-tune them.
I can’t claim to understand fully what I’m doing, although I think I grasp it intuitively. When I read what expert researchers have written on some of the simpler methods I decided aren’t good enough, the math quickly overwhelms me, so I imagine that I am working with much more complex math in my new algorithms. (I also assume that my new algorithms are also reinvented wheels, and I will probably find out what they’re really called at some future date.)
I’ll present some of my work at Percona Live in a few weeks.