Searching & clustering traces

[jonnor]

When capturing traces from production environment with lots of processing, one often encounters scenarios where many failures are in fact caused by same bug/issue. And that small changes in inputs often are difference between triggering an issue or not, and that such differences can say a lot about the nature of the bug.

Interesting questions include:

• How many distinct failure classes exists, and how many of each are there
• What are the boundary conditions between fail and success

With the ability to upload traces to a service (#2), it becomes interesting to have cloud-based tools which help answer these problem at scale. Performance of making the needed analysis - both how quick to get results, and how much it costs, are key considerations.

References

• https://en.wikipedia.org/wiki/Cluster_analysis
• http://www.cs.cmu.edu/~ninamf/papers/clustering-bbv-long.pdf

[jonnor]

Need to do a review of exiting practices, tools and services that do such things. Some domains that come to mind are:

• Free text search/indexing
• Crashdump / stacktrace collecting services