Data mining logs to locate a bug

Question

I'm working on a data distribution application which receives data from a source and distributes that data to multiple target application. After successfully distributing several messages each second for 8 days, it missed a single message and did not deliver it properly to the clients.

As I was looking at the logs I tried to find something there that was special for the time the miss happend - either in the data, its rate or some other condition but couldn't find anything.

Is there any data mining technique I can use to identify how that specific event differs from other events?

Answer 1

You may look at outlier detection. A basic technique is to plot several variables for lots of messages and see whether this message literally stands out. A way to visualize more than two dimensions simultaneously is parallel coordinates. Maybe you should also look at, say, half-hour bins, calculate their statistics and see whether there was a problem around this message.

Answer 2

I don't think data mining is the right tool here.

I would add some systematic logging (at interfaces, hard to deduce internal state changes, and in your case why the decision to distribute or not was taken) and try to reproduce the problem, then figure out from the logs exactly what went wrong. Though it might be hard with millions of messages and a rarely occurring bug like this.

Answer 3

Sorry, I fail to see how this is anything but a misunderstanding of what data mining is.

You already know the troubled message, and hence you know it's an 'outlier.' So what are you 'mining' for?