# codeswap: an experiment idea to compare coding style efficiencies

i am really curious how effective my "avoid creating types" approach would be compared to the object oriented approach. i am thinking of creating an experiment for it. not that i would ever conduct such an experiment but thinking about it is fun. i am not even sure where would i get people for this experiment but suppose that problem is solved. now imagine i have a room full workstations next to each other like in a typical open plan office. the experiment would involve having the participants solve a series of related programming challenges on these machines under a time constraint. then they would need to swap their solutions and continue working on other people's code. the swap also changes the problem the participants are supposed to solve. there are two problems, the swap is done a way that it changes the problem for everyone. to make sure there is no language barrier, i would ask each participant to use the same language. something simple, e.g. c, go or python. i would ask the participants to brush up their language skills before coming to my test. i would also make the environment equally bad for each participant so that the environment has very little effect on the results. to be specific: all i would give them is just a terminal that is split into two windows: left side runs the nano text editor, right side runs a terminal app that shows how correct their solution is. every time they hit save in nano, it updates the right hand side with the results. i would split the experiment itself into the following phases:

• intro phase: give participants their problems and have them read a motivational text how to approach the problems.
• design phase (30m): the participants can write code but they get no feedback on the right hand side yet. they are supposed to architect their solution during this time. nearing the end they are also expected to write a feedback how hard they thought the problem was or their general thoughts about the experiment. they can just write this feedback directly into their source code as free form text at the bottom, i will strip this out before compiling or swapping the code.
• implementation phase (30m): at this point the right hand window wakes up and starts giving feedback. if they worked well, the participants should be able to easily make their code compile and solve the subproblems correctly (the ones they have already implemented) and then just continue solving more subproblems. nearing the end they are expected to give feedback to the study at this point too. however at this point they should also leave some pointers what was done and what the next steps are for implementing more features. this should be aimed at the next person.
• continuation phase (60m): now the system swaps around the code. the students have to work with other people's code. they have to get familiar with it and continue working on implementing the missing features. the swapping phase also changes the problem so the students also need to get familiar with a different problem. nearing the end they should write more feedback.

i am not sure how to implement the swapping phase yet: i can either ask the participants to swap desks, or i can make the computers shuffle the code around. either one works. if the latter, i need to be careful to prevent participants from saving utility functions somewhere for reuse. they should completely lose everything they have written so far.

the two problems would be very similar when reduced to the core but they would be phrased differently and put into different context so that it is not obvious that it is the same thing. but at the same time the two problems must be different enough so that the participant cannot just simply delete everything and rewrite stuff from memory.

the motivational text would have 2 dimensions, totaling in 4 different texts:

• one half says that having comments is an indicator of poor code, participants should design their code in a way that does not require comments. the other half says that comments really humanize the code, people should write a lot of comments to make the intent clear.
• in the other dimension one half says that abstraction is what keeps the code simple and straightforward. make the structure of the code similar to the structure of the problem. the other half suggests to avoid abstraction: try avoiding creating new types, try sticking to arrays and long functions.

after the experiment is done, i would ask them to fill a survey. for each phase it would ask the following binary questions:

• was it clear what you needed to do?
• did you find this task fun?
• did you feel productive?
• was it easy to work with the code?
• did complexity unnecessarily hindered you?
• do you feel you could have achieved more in a better environment?

my hunch is that in the design/impl phase the nocomments+abstraction group will feel the most fun, ease and productivity but the comments+noabstraction group will get more stuff done. in the continuation phase:

• nocomments+abstraction continues on nocomments+abstraction code: not fun, hard, unproductive.
• nocomments+abstraction continue on comments+noabstraction code: not fun, easy, productive.
• comments+noabstraction continue on nocomments+abstraction code: not fun, hard, unproductive.
• comments+noabstraction continue on comments+noabstraction code: not fun, easy, productive.

in other words: the continuation phase is not fun since people do not really like to work on other people's code. working on non-abstract code will be much easier and people will feel more productive afterwards. the "comment+noabstraction continues on comments+noabstractions" group will have an advantage though: if they followed the recommendation, they should have come up with roughly the same structure so they should be more familiar with the new codebase.

and if it turns out that is the case then i will be vindicated. most software is just a different configuration of the same underlying concepts. there is no need to abstract away all this but rather what is needed is learning to recognize patterns. the professional chess example is very apt: pro chess players can instantly recognize very complex state in the middle of a pro match. even though they might even give some states a name (an abstraction), it is not the name that helps them recognize the state but rather the fact that they worked with so many different chess states so long. it is the experience that matters rather than ensuring every little concept has its own name. in fact, take two random pro chess players: if they talk different languages they will not be able to talk with each other much. but show both of them the same chess board and it is very likely that both of them will make the same next step. this next step is obvious from the abstraction-free chessboard. yet if they cannot see the chessboard, they will pretty much not be able to agree what should be the next step solely because they do not know what the other person is talking about.

when working with code, all we can do is to project the code's concepts into concepts we already know. for instance as part of my sysadmin work i often project the software into the realm of linux system calls. i am very familiar with those and as long as i can understand the code in terms of syscalls, i can pretty much have a rough idea what it is trying to achieve and how. a code feels complex when i cannot do this projection. so my recommendation is to stay on this level, avoid creating concepts that have unclear projections onto the well known concepts. and i am not saying to completely avoid creating new concepts. rather only create a new concept if you have a handful of usecases (at least three): then i can see the concept from various angles and i can learn why is it an useful concept. otherwise it is just a like a useless signpost saying "this is a signpost".

anyways.

as for the problem: i am not sure about this yet. it needs a theme and a bunch of small and easy problems related to that theme so the participants can solve some of them. maybe something like this: you are managing a garden. your task is to read the garden description and help the farmers answer a few questions:

• what is the area of the garden?
• how many types of vegetables are there?
• which vegetable consumes the most area?
• can you render a map of the garden? (just on the console, not a picture)
• given the work needed for each vegetable, how much work does the farm need?
• how expensive is the garden to maintain?
• how much money can the farmer make of the garden?
• suggest a better, more money making plantation for next year.
• which two plantations are the furthest away?
• and a very hard one for the ultra productive people so that they do not get bored: can you draw a road network to connect the plantations? make sure everything is easily reachable but at the same time try to minimize the number of road cells you need.

obviously i need to think more about this part but that is the rough idea. the other problem could be very similar but it would be zoo themed. you would have animals and zookeepers instead. or maybe school themed where you would have teachers, students and grades. dunno, this is just an implementation detail.

thinking about this even gives me an idea on how to approach learning the ways of software development if you have access to a buddy: both should start a project each but nearing the end have the other person finish it. iterate this (while occassionally changing buddies) until you feel confident in coding. every time you do a swap, there is a lot of context that needs exchanging. but if you do this often enough, over time it will be clear what sort of documentation or coding styles you need to write or use to make your code easy to pick up. i wish more people would be better experienced in this area of software development. oh well. might be a silly idea but still interesting to me.

maybe if i were a university professor with lots of time on his hand, i could conduct such experiments on my students. hmm.

published on 2019-03-03

to the frontpage