Tools for metrics analysis and fault detection. It's currently experimental only. Have a look at the [wiki] (https://github.com/adrien-mogenet/opengossip/wiki) for current work's output.

There are sevenral kind of metrics that you would want to track:

• Metrics that should always be evolving and never be constant
• Metrics that are highly related to each other
• Metrics that should be perfectly predictable

Of course, putting this or that metric into this or that model depends on your platform and programs that are running on it. OpenGossip has been designed as a toolbox that you might customize to fit your own needs.

OpenGossip is mainly written in Python so far, and Octave is used when prototyping different algorithms. For the best use of the current source code, you might want to install:

• python >= 2.7
• numpy
• pypy >= 1.9
• numpypy
• gnuplot >= 4.2
• octave >= 3.7

Pypy and numpypy are both not strictly required, but are speeding up the execution (at least x2 faster).

OpenGossip is gathering several models to play with for metrics analysis. It extracts features for further analysis. For example, when dealing with time series, it will probably work with such features (among many others):

• mean
• standard deviation
• slope
• entropy

OpenGossip is learning a lot from these characteristics and will be able to classify your data as normal or as anomaly. Currently, learning model(s) is/are not yet established, and the "perfect" model is still pending... Sometimes naive threshold-heuristics will get the best F-Score.

Here is the graphic representing a load average (based on 1 last minute) :

And now, how to automatically detect spikes ? Using classification algorithms mentionned above (here a hierarchical clustering technic) OpenGossip is able to detect abrupt changes, strange behaviors or related events.

Here is a Multivariate Gaussian implementation, in Clojure (using Incanter):

It is still in development, and even at an early stage. It currently mostly uses naive implementation (sometimes very high complexity) and is continuously developed to determine best way to find anomalies among your metrics.