Code Review - a Needed Habit in Science

This post originally appeared on the Software Carpentry website.

More than a year ago, Marian Petre and Greg Wilson wrote an article about reviewing code, but I was not aware of it until last week when Greg talked about their findings in a workshop in London (though it had been mentioned here before). I have been reviewing code for years, however I have not realised that I was doing so till GitHub made it easy. Reading code from others feels the same as proofreading a Spanish text (or any other text), where you have to pay attention to the orthography, the grammar and whether it says what it's meant to say (or it does what it's meant to do).

Since I started my PhD I've organised programming clubs in the institutes where I've been as a way of sharing knowledge between the colleagues. But in these last 10 years I've just done a "code review" session twice. The first one involved external people (from industry) reading scientific code, whereas the second time (last week) were done between the researchers.

The first session was as a test run for a project that a friend and I had in mind: "Help a Scientist day". Our idea was to get software developers to look into code written by scientists. But before we create a huge event we run a test inviting some developers we knew for an evening with free pizza to see some piece of code written by PhD students in my group. We had three codes to review, they were written in IDL and you could almost notice the learning curve of the students through out the code. Two of these were a single file with almost 2000 lines of code, without functions and very little documentation. These codes were now being required to be run in a pipeline of a NASA mission, something the students had never expected during the three years before. The codes were made available a week or so in advanced, but the authors received very little feedback in the code itself. The session was converted more into a discussion on good practices than a review per-se. I believe many of the students present there (beside the authors) learnt a lot from it. A very important question from the developers was "how much time we spend coding", and our answer was something like 80% of our time. They were struck that we didn't talk and discuss more about coding. Leaving the choose of language used a side, we should be experts in the one we work. Though many of us believe we are, we are not and we get surprised when we read some details we were not aware of after 10 years using it. Other outcomes included things like the use of version control, thoughtful variable names, and read others code. Little time after that I moved country for a new postdoc position and our big event of "help a scientist day" never happened.

Four years later (i.e., last week), I proposed an exercise to the attendees of the programming club I do at work. The rules were, find some piece of code you wrote a while ago and make a pull-request to a repository I set up for such exercise. The first rule was done for two reasons: to avoid them to write something for the session knowing they were going to be read by others; and to see how much our present-I understands what our past-I wrote. We were just six people, and we've got three pieces of code in IDL and three in Python. During the one hour session we had I started with a very quick introduction on how GitHub's pull-request commenting tool works. I didn't explain what they have to look for, I just told them to comment into anything they didn't understand. So we paired us into languages for making the review as if pair-programming, to bring up discussions between the pairs... and to "make matters worse" I then asked the IDL teams to review a code in Python and vice versa. At first they were a bit lost on the language but it worked great to highlight things that were not documented properly or style. Like why the 14th element in this Python line:

pos_x = float(tstep.split()[14])

int_upper = interpol(  radius_tot, density_pure_model[*,39], density_upper )

We also found factors added or subtracted to variables which the authors didn't remember why was done, or easily identified "copy-paste"-ed blocks. Then we look all together to the comments each pair have done and continued the session by each team reviewing another piece of code but this time in the language they knew. In this case the comments focused more on how to do things differently to help readability or to make such algorithms more general. For example how to avoid repetition by using a combination of lists and for loops or parsing arguments from the function itself. All of them really enjoyed the session -or so they told me- and learnt not just to review but also realised how hard is sometimes for our future-self to understand what we wrote.

In summary I don't think I can add much more to what is in Marian and Greg's paper but to emphasise the need to do such activity in our research groups, to raise its importance as when we review papers as these codes are the key for science progress.

I think it's worth mention different justifications I've heard during the years to don't share code like:

• it's not well commented and you won't understand it;
• I will show the code when the paper is published;
• The code has some drawn-backs and if I leave it there to people to use they will publish wrong things without understanding it.;
• I want people to contact me to run it, so I can be in their publications.

I believe that if more researchers get used to review code, many of these comments will disappear (maybe not the last one; though people may not believe them). What do you think? How can we increase such awareness? what things have you tried out? which excuses have you found and how you "fight them back"?