A lot of changes are happening to science as I write this. Crowdsourcing, open reviews, citeable data, new ways to measure contributions, automatically tracking the provenance of every calculation—each has emerged from its chrysalis and is just waiting for its wings to dry so that it can take flight.
Our job is to give scientists the skills they need to nurture these ideas. For the last post in this series, I'd therefore like to talk about what victory will look like—i.e., how scientists' lives will be different if and when the things we teach become as routine as doing statistics. My vision is simple:
Scientists won't submit, publish, and download papers. They will fork and merge projects.
The single most important thing we teach scientists is version control. We tell them it's for record-keeping, but it's also the underpinning for a style of work that open source developers think of as normal, and the rest of the world finds extraordinary. Do you like my code? Do you think you can improve on it? Great! Fork my repository, build your stuff on top of mine, and then give me your improvements to merge into the master copy so that the next person has an even better starting point. Firefox and Wikipedia are two proofs among many that this can scale to thousands upon thousands of contributors. There's no reason it couldn't be the "new normal" for science once version control and a few other things are routine parts of scientific life.
What exactly are those "other things"? The first is task automation: everything someone might have to re-do must be captured in a re-runnable way. This can be a shell script, a Makefile, a snippet of Python or R, or something as yet undevised. What matters is that without this, people have to invest hours or weeks of forensic investigation before they can use someone else's work, which in practice means they won't.
The second thing is comprehensive testing. This isn't just a special case of task automation, although tests that aren't automated might as well not exist as far as people forking a piece of work are concerned. It isn't primarily about ensuring that software is doing what it's supposed to, either. Instead, as we teach people in our boot camps, the primary purpose of testing is to define what the author thinks "correct" actually means. If you tell me that the eigenvalues of a matrix are close to zero, I don't know what you actually mean, any more than I know what you mean if you say that the heights and weights of a sample population are strongly correlated. If your tests check that all the eigenvalues are less than 0.0001, on the other hand, we now have something concrete to agree or disagree about.
And yes, it's often hard to test scientific code, and it's also often not even appropriate—just think of the twenty lines of R that turn a bunch of CSV files into a line graph. But that brings us to the third of our "other things", which includes things like structured program design and data management. For lack of a better name, I'll call it computational tidiness; as with task automation, it's what allows other people to pick up a piece of work and build on it without superhuman effort. It's also what allows people to review work: if they can't find things easily, or make sense of them, they won't bother.
We teach the Unix shell, Python (or R), Git, xUnit testing libraries, SQL, regular expressions, and sundry other things because big ideas aren't meaningful in the abstract. Along the way we hope to convey a deeper understanding of seven big ideas that I believe can only come from getting your hands dirty:
If, ten years from now, the average scientist understands these as well as she understands significance, correlation coefficients, and p-values, I for one will think that we've won.comments powered by Disqus