Official Blog
learning data science
+1

# Using Data Science to Explore Software Development

What can data science mean for software development? In this blog post, you'll discover some interesting case studies of data science in software engineering!

Many software developers are learning data science to analyze their customers' data. What a growing number are realizing is that they can use those same techniques to answer their own questions, such as:

• When will this project be ready to ship?
• Which components of our application most need to be tested?
• Who should fix this bug?
• What parts of my API do people find hardest to use?

Over the past 15 years, there has been an explosion of empirical research in software engineering to explore these questions, fuelled in part by the availability of data from sites like GitHub and Stack Overflow. This blog post presents a few representative results, then explores the methods behind three in more detail.

If you'd like to know more about how to use data science to get insights like these into your own projects, please let me know.

## Data Science and Software Engineering: Examples

Let's start with a finding that affects everyone doing data science at scale: Yuan et al's discovery that simple testing can prevent most critical failures in distributed data-intensive systems. They examined failures on Cassandra, Hadoop MapReduce, and similar systems and found that:

1. Almost all failures required 3 or fewer compute nodes to reproduce. That is, you usually don't need a cluster to debug a cluster.
2. Error logs usually contained enough data to allow reproduction.
3. The majority of catastrophic failures could easily have been prevented by performing simple testing on error handling code.

The last point is the most unexpected. While professionals usually test that their analysis code works when things are going well, they rarely test that it does the right thing when something goes wrong. Adding just a few such tests during development would prevent a lot of pain downstream.

Another large-scale result comes from the work of Altadmri and Brown, who looked at 37 million attempts to compile programs by high schoolers in the United Kingdom. They then asked teachers what their students' most common mistakes were, and discovered that:

1. Teachers didn't agree with each other about what mistakes students were most likely to make.
2. More importantly, teachers' predictions had only weak correlation with what students actually got wrong.

Of course, data mining isn't the only way to generate useful evidence-based insights in programming. In 2013, Andreas Stefik's team published the second in a series of studies of whether some languages are easier to learn than others. As a baseline, they included a made-up language in their study whose keywords were generated at random.

To their surprise, curly-brace languages like Java and Perl were just as hard for novices to learn to recognize as a randomly-generated language. Python and Ruby were both significantly easier to learn, and Quorum, which A/B tests the syntax of every new feature before adding it, was easier still. Andreas discusses this result, and the reasons why programming language designers pay so little attention to usability, in this entertaining podcast.

## Identifying Security Bug Reports

As an example of how these studies are done, Fayola Peters at LERO used text-based prediction models to find reports of security bugs in the bug trackers of large systems so that work on them could be prioritized. In one study, she used data from the Chromium, Wicket, Ambari, Camel, and Derby projects; these contain a total of 45,940 bug reports, but only 0.8% are security bug reports.

Some security bugs are manually labelled, and these can be used to train classifiers to detect unlabelled ones. Techniques like Naïve Bayes and Random Forests are a starting point, but the efficiency of generic models can often be improved via filtering of specific keywords. The most important check is how well models built from projects with labelled security bugs can find similar bugs in other projects. The results were encouraging: Peters found that her classifiers worked well enough to be useful in practice.

## Will My Patch Be Accepted?

As a second example, companies want their modifications be included in projects in order to avoid having to maintain the code themselves, and volunteers want an indication whether their patch stands a chance as well. Since there is often a long time between patch submission and acceptance, Bram Adams and his group at the Polytechnique Montréal have been building models to predict which ones will be accepted.

Like many data science projects, this one starts by getting and cleaning data from multiple sources, including Git repositories and Gerrit code reviews. However, for projects that use email-based review, like the Linux kernel, gathering data means scraping mailing list archives and using heuristics like the checksums of patches or set-based intersection of changed lines to match patches to conversations. Even with the help of tools like GrimoireLab, there will always be patches of which multiple versions were reviewed, and patches split into multiple parts that were reviewed separately. As in all data science, it is up to the analyst to build a model, and the assumptions in it will have a powerful impact on the final conclusions.

The second step is to decide what variables to examine and what metric to use for each, such as:

• patch quality
• the experience of the patch developer
• the review process (for example, review comments and number of reviewers)
• review quality (for example, degree of detail of reviews)
• review time
• interaction of patch author and reviewers (for example, friendly versus angry)

Just this half dozen might use everything from survival statistics to sentiment analysis. Once deciding what to measure, you can apply the data science tools that DataCamp's courses teach. Since your real goal is to predict acceptance of future patches, the most straightforward approach is logistic regression, but there are many others.

Finally, you can measure variable importance to determine the impact of each variable on patch acceptance. This will necessarily be partly qualitative, since your goal is to identify what developers can do to improve the odds of their work making it into the product. As always, you have to be careful about confusing correlation with causation, but even a few simple recommendations (like "DON'T USE ALL CAPS IN YOUR COMMIT MESSAGES") can make a real difference.

## Finding Help

The biggest change in how programmers work over the last 20 years hasn't been in the languages they use: it has been the near-universal reliance on Stack Overflow for asking questions and getting answers. Christoph Treude at the University of Adelaide and his collaborators have been developing methods to make sites like this more useful.

Their starting point is the Stack Overflow data dump. After doing some exploratory statistics to look at the number of words and sentences, most common words, term frequency–inverse document frequency (TF-IDF), and so on, the next step is to see which words in the Stack Overflow threads co-occur with higher votes, higher acceptance rates, and higher view counts.

Since software documentation contains many incomplete sentences and code elements, off-the-shelf natural language processing libraries often get things wrong. Software documentation written in languages other than English is even more difficult to parse since it still tends to use English for technical terminology. That is, it mixes two natural languages and code elements. Ad hoc heuristics and more advanced modeling techniques have to be used together to handle cases like this.

Putting the pieces together makes it possible to build a tool that takes the name of a Python module as input and produces meaningful sentences about this module from the Stack Overflow threads. Going further, Treude and his collaborators were able to use grammatical dependencies between words to automatically find software documentation that explains how to perform a task, and then to automatically identify code snippets that can accomplish this task.

## Conclusion

All too often, widely-held truths about software development are based on strong opinions and loud voices rather than evidence. As described at the outset, that is changing as hundreds of high-quality studies appear every year to support some beliefs, such as "code review really is the best way to find bugs", and challenge others, like "test-driven development isn't as effective as some people believe, and goto statements aren't really harmful".

"Engineering" has been defined as "the application of the scientific method to create useful things". If that's true, then software development might finally, thanks to data science, be on its way to becoming a real engineering discipline. If you would like to see courses to show you how to apply data science to software development, please let us know.