Blog post from Coscale on the use of anomaly detection for web performance monitoring:
Anomaly detection (or outlier detection) builds on elements from statistics and signal processing, as well as from machine learning and AI, which are surrounded by a lot of hype. As often with a hype, there are camps of believers and non-believers, and for anomaly detection this is no exception. And for the right reasons. Anomaly detection has sometimes been overpromised as a panacea that works out of the box and solves all your problems automagically. The reality is that anomaly detection is a really hard problem to solve, in particular for web performance monitoring, as also mentioned in a previous post on the topic. There are many reasons for this: the large number of metrics, their different distributions, their interdependencies, the subjective qualification of an anomaly, etc. Techniques that have worked in other fields can not just be transplanted, which we have found out the hard way.
We need anomaly detection for networking monitoring. We manage a distributed system of autonomous elements where the operation, performance and stability is determined by external factors that are out of our operational control.
Operations doesn’t need to know when its working, we need to know when its not working the way it used to. Thats anomaly detection.
I want it.
PS: If you have it I don’t know about it. Please get in touch