Data Science Digest: Source Code Metrics That Matter
Data Scientists have spent decades researching ways to better understand what’s “good” code and what isn’t. While software quality is an interesting area to explore, is it worth the effort? After all, it’s common for developers to disagree on how to write high-quality code to achieve the same outcome.
Dr. Maciej Gryka, Sr. Data Science Manager and self-proclaimed “Data Nerd”, recently sat down for a webinar to discuss his research learnings in Data Science and Machine Learning on the topic of software quality. From source code metrics that matter to where the research is going with Machine Learning, these are the key takeaways from the discussion:
Software Quality Is So Much More Than Functionality
You wouldn’t necessarily call something that works well high quality. You also wouldn’t call something that looks nice, but doesn’t work well, high quality. You need both.”
While Maciej primarily covered source code metrics, he emphasized they should not be the only factor taken into consideration when measuring software quality.
The software development community often focuses on functionality, or how well the primary purpose is served, when measuring software quality. But like any non-software product customers use and love, software needs to be both functional and delightful for it to be high quality.
Consider these questions:
- Is your software solving the right problem?
- Is it solving it in the best way?
- Are your users happy while using the software?
- Are your developers happy while working on the software?
As we can see, there are so many ways to measure software quality beyond functionality. The key to amazingly high quality software is in both functionality and delight.
Source Code Metrics That Matter and How To Use Them
The software development and research community typically looks to a few key source code metrics to measure software quality. While we can extract numbers from these source code metrics, like complexity and hierarchy, how do we know which numbers are desirable?
This is where historical data and Machine Learning can help: we can approach it as a supervised learning problem, and train a model on which code metrics are indicative of potential bugs. We can understand how to do this from the most well-known source code metrics:
- Number of Bugs Found: The most-obvious and popular way of measuring code quality. Note that tracking when in a feature’s lifecycle a bug is found is important; as a bug found early in development has a different impact on quality than one found later in production.
- Number of Code Changes: With every change, the chance of introducing a defect increases, so roughly speaking, the more changes there are to a piece of code the more likely it is to have defects.
- Proxy; Complexity & Churn: Churn is measured as how many changes there are to a file or function. The higher the churn, the more often the file or function is changed. Complexity refers to how complex code is. We can look at the relationship of these source code metrics to measure quality:
While it’s unlikely there will ever be a single number to represent all aspects of quality for a given code base, source code metrics are interesting to explore because they can show us what kinds of code constructs have historically led to defects.
Avoid Using Code Quality Metrics to Measure Developer Productivity
It can be tempting to use code quality metrics to measure developer productivity. Do not do this, it’s dangerous!
Evaluating developers based on code quality metrics is a sure-fire way to get them focusing on maximizing metrics, rather than on shipping high quality code. All known metrics are imperfect and gameable, so make sure you do not create misaligned incentives in your team’s software development culture.
Learn How Machine Learning is Changing Software Quality Measurement
Testing every piece of code before shipping can be a lot to ask. Sometimes it’s not possible, or is very expensive, so we want to focus on the most important parts. Ideally, we could predict the number of bugs in a piece of code before it hits production.
While commonly-used source code metrics are reasonably good at measuring software quality after code hits production, they do not help prevent bugs from getting into customer’s hands. That’s where Machine Learning comes into play.
To learn more about how Machine Learning is changing the field of software quality measurement, listen on-demand to “Data Science Digest: Source Code Metrics that Matter“. Our ‘Resources‘ page is also available to you for more content on how to level-up your QA strategy and achieve higher quality products.