Lesson learned after programming full-time for a year.

It’s been more than a year now since I have been programming full-time. And I would like to reflect on the kind of problems that I encounter most often.

I don’t actually want to talk about specific problems that I have solved rather about one meta-problem that have occurred time and time again.

The most common meta-problem is to get the code that was written to do almost what I want it to do but not quite to do exactly what I want it to do.

Whether it is a change in requirements, or me developing some component that relies on the existing infrastructure, or me spotting that two independent components that are doing exactly the same thing and should actually be refactored to be unified as to avoid duplication: there is always something out there in the codebase that is almost there but not quite. So 95% of the code is already there, but the last 5% often takes days of work to write and plug in.

And if you read “Code Complete: A Practical Handbook of Software Construction”, top voted pick of “What is the single most influential book every programmer should read?” on StackOverflow, you will know that a big chunk of the book advocates planning ahead exactly so that this sort of scenarios don’t happen often: “If you can avoid writing 70 percent of your code by planning ahead, do it!”, says the book. More quotes from the book on the subject: “It takes only a few large programs to learn that you can avoid a lot of stress by planning ahead.”, “From a technical point of view, planning means understanding what you want to build so that you don’t waste money building the wrong thing.”.

I, however, have grown to be pessimistic about the value of such planning. Or at least elaborate planning. My planning now solely consists of decoupling as much as possible in an anticipation of unforeseen changes (the book will also call this strategy planning, however it appears cheating to me: might as well just say “write good software” at this stage).

Seriously though, if you can foresee exactly which components will be changed months in the future - I think you are wasting your talents writing software. You should do something more significant with such forecasting abilities. Every time I modified either my own code or somebody else’s I always know the reason why it was written the way it is written. And it was always done so for good reasons: it was written with a goal X in mind, not Y. To be slightly more concrete, this is a somewhat common situation.

func validateAndPostEvent(event *Event, endPoint *Url) {
// validates the event against a schema
// ...
// posts it
}

Validation and posting is coupled in the above situation because there’s was no other reason to validate the event separately from posting at the time when the code was written. And now that I am writing say, a static schema verifier, I need that validation part of the code, but I need to figure out how to decouple it away from posting.

Another occurrence is:

def score_bunch_of_rows(csv_in, csv_out, model):
# read csv
# process csv
# score using model.score
# write csv out

However this time I need to score using model.contribtutions function, so I want to reuse most of the code above, but change just 1 line in the middle. However read_csv, process_csv, model.score & write csv are all coupled in the above function. The reason it is all coupled because the author, me, didn’t think at the time that I would need to call .contributions in the above function.

These are toy examples that make it appear that decoupling from the start was a negligible amount of effort, but as I already mentioned, it often takes me days to tear certain functions apart: they were coupled for good reasons at the time that are hard to communicate. Real world software is complex.


Like I already mentioned above, my software writing strategy has steered towards: try to decouple code components as much as possible right from the start even if it’s not clear to you now how these functions could potentially be reused right now. Concretely it means: if your function takes a struct but only uses 4 fields of the struct, then create a function that operates on 4 fields only and another function that takes a struct but calls the first function instead with the required fields. If your function requires an object with .predict method, you might as well just take the .predict method rather than an instance of an object as function argument. This way the callee doesn’t have to mock the interface of the entire class just to so that you get to call .predict.

OOP-wise it means having classes as lean as possible. If you have, for instance, the following abstract class:

class PersonRecord(abs.ABC):
def __init__(firstName, lastName, email, phoneNumber):
pass

def phoneCountry(self):
# compute phoneCountry from self.phoneNumber

@abc.abstractmethod
def fullName(self):
# insist derived classes implement this

This means you now can’t call phoneCountry function on a free text telephone number to find what country the phone number is from without creating an instance with irrelevant details such as firstName, lastName. .phoneCountry and names are coupled even though they should not be. In addition, you have to hunt around for a concrete implementations that might have their own idiosyncratic requirements in order to call our .phoneCountry method. Not cool. (Another way to think about it: every time you add a field to a class you create an implicit dependency between all methods of that class and the field! Unless all of the methods will make use of the field, you just created an unnecessary dependency.)

And this is the meta-lesson that I learned. Steer towards fewer dependencies between functions and classes. Only require what you need in a function. Only put things into a struct when you are very sure these things will always be used together. Avoid creating classes if you can. Tada.

P.S. I suspect none of this is new, however I don’t know the name for this technique.