80% of machine learning is said to be data wrangling. Is all this wasted effort? Hardly - often the journey really is its own reward.

In this talk, we'll briefly describe a machine learning project that predicts the outcome of test cases in a large-scale software development cycle. We'll then show what we gained from collecting the necessary data and how these insights can have lasting impact on the day-to-day work of developers, testers and architects. This includes a quick answer to the well-known question: Whose defect is it anyway?

Target Audience: Developers, Testers, Architects
Prerequisites: Basic knowledge of software development and testing and an interest in data analytics
Level: Basic

Extended Abstract:
Data science folk wisdom holds that at least 80% of machine learning consists of data wrangling, i.e. finding, integrating, annotating, and cleaning the necessary data.

While sometimes viewed as less "glorious" than the deployment of powerful models, this journey has its own rewards as well.

Benefits may sometimes be somewhat expectable, if still non-trivial, like data cleaning potentially exposing errors in underlying data bases.

In other cases, though, there may be some low-hanging fruit a casual glance might miss even though they are indeed rewarding.

Data integration often reveals opportunities for statistical analyses that are relatively simple, but may still have high impact.

In this talk, we'll start at the destination: the result of a machine learning project where we successfully predicted test results from code changes.

A necessary task was the integration of several data sources from the full software development cycle - from code to testing to release in a large industry project with more than 500 developers.
Naturally, this required all typical steps in the data science cycle: building up domain knowledge and problem understanding, both semantic and technical data integration, data base architecture and administration, machine learning feature design, model training and evaluation, and communicating results to stakeholders.

These steps yielded several benefits, on which we will focus in our talk.

Among others these include data quality insights (e.g. about actual "back to the future" timestamps), and new analyses which were made possible by a unified view of the data (e.g. survival analysis of defects).

Last but not least, we demonstrate a simple answer to a well-known question, especially in larger software development contexts: Whose defect is it anyway - how can we avoid assigning defects to teams that have nothing to do with them?

Thanks to the integrated information sources from systems concerned with version control, test result logging, and defect management, we are able to support the claims made in this talk with statistics taken from 6 years of real-world data.