Code boilerplate: Is it always bad?
medium.com
medium.com
Generalisation is always[tm] a bad idea. Taken to illogical extremes, at the other end of the spectrum we have the monstrocities of old Perl. After all, with Perl it was often said that you could fit most solutions to a oneliner.
In fact, I would go as far as to claim that using Python for n-tuple iterators inside branched comprehensions are getting awfully close to the worst aspects of what Perl scripts were commonly criticised for.
1. RESTful routes leading to implied controller actions that render a view without any controller declaration at all. I've seen more than one codebase with routes that could be invoked by users despite being never linked in production and having no test coverage. It's basically undocumented code due to a total absence of declaration, and a case of needing to know the implied framework behaviour to understand a program.
2. I have an issue with ActiveRecord's schema behaviour. By avoiding attribute declaration in models and just defining ORM methods by asking the database for column names, it only takes one poorly handled migration or a DBA oops to lead to some very strange production behaviour. Situation is only partially offset via the (new) attributes API and/or a schema cache dump during CI but one can never fully do away with the hazard. Philosophically speaking, sometimes Rails thinks it is dictating the database schema, sometimes Rails accepts what the database tells it, and out-of-the-box rails has no model boilerplate that bridges the gap to guarantee a specific DB schema.
In both cases I add my own safeguards for production environments, but that's a matter of programmer discipline rather than mandatory declaration.
Boilerplate is usually maximized in projects which try to give users a huge amount of flexibility that they don't need or want.
This is partly why minimizing boilerplate is so hard - one thing that continually surprises me is what users want flexibility over and what they don't. I've lost count of the number of times I've given users flexibility to do something they don't give a shit about and not given them the flexibility to do something they really wanted.
That's a great way to put it. I'll steal this one.
I tend to use this pattern quite a lot:
#region Find the first Foo with Bar
var foo =
foos.Where(f => f.Bar)
.Distinct()
.OrderBy(f => f.Baz)
.FirstOrDefault();
#endregion
I already pondered if it wouldn't be possible to merge multiple "folding strategies" if one would build some kind of interface with which multiple routines could basically communicate folds in a buffer.
Say you have one routine scanning the syntax of the according language, one looking for folding markers and one for manual (Vim-style) folds. I guess it technically should be trivial to use all of them provided they are nested and don't intersect in weird ways (the latter could be resolved with some simple priorities).
Simply put, knowing how advanced org-mode can get with folds, this is likely trivial to do. (If you haven't played with it, I'd recommend taking a look. I don't have perfect examples, but http://taeric.github.io/DancingLinks.org is an example. Open that in emacs with org-mode and you will see that all of the source blocks can be folded together. For me, the source blocks are correctly colored, as well.)
Quickly searching, I found https://www.emacswiki.org/emacs/FoldingMode, which seems to be a large start to this. There were also some other ideas that seem to get close.
The basic idea I had was to just make a few functions that would:
- Fold from current location to next #endregion
- List all region/endregion's in the current buffer
- Fold all region/endregion's in the current buffer
Edit: It occurs to me that I didn't show how to actually hide text, which I think is your question. Fastest way I can see is this (run this in your scratch buffer to see it in action):
(let ((o (make-overlay 5 120)))
(overlay-put o 'display "---hidden---"))
(let ((o (overlays-at 5)))
(mapcar 'delete-overlay o))
// Find the first foo with bar foos.map(f => f.bar).uniq().orderBy(f => f.baz)[0]
does the same basic thing, the ability to collapse and display intent is incredibly rad for lack of a better word
I agree with you, but seeing the number of proponents of Enterprise Java there are, it's clear that there are a significant number of programmers who prefer verbosity and extra abstractions --- although IMHO not necessarily because it's actually better.
In fact a trend I'm seeing with a lot of programming languages is more verbosity and unnecessary abstraction, leading me to wonder whether programmers actually know the languages they're using or are barely scraping by with the very minimal basics (and in the case of some languages, a bit of OOP cargo-cult dogmatism.)
Although I'm not really familiar with Python, and the order of the clauses its conditional expression syntax is unusual and a little surprising, I think that second line probably took longer to write than it did for me to read.
Even Enterprise Java wasn't verbose for the sake of it (I say "wasn't" because modern enterprise Java can be more like modern functional style than Go).
It was more verbose because it added tons of options and flexibility while keeping with the OO style, so everything was pluggable and configurable and factorable. But that's something else than Go style verbosity.
Say, with Scala implicit parameters, you get to remove code from call-site. Does it make the code better? Depends on the case, in my opinion. So I don't get the point to say some people categorically want verbosity. They may prefer explicitness and unsurprising code which may require verbosity. Or perhaps the removal of verbosity requires code that allocates more memory -- in that regard Kotlin seems to be a good citizen. It has ways to make code concise without making the byte code entirely different.
That's the point. Reading is far more important.
It is more that they have good abstract thinking, so abstraction is not an issue for them. Abstractions are issue only for some people.
> I think that second line probably took longer to write than it did for me to read.
If it takes more time to read then write, then it is not done yet needs to be refactored. Good code is as obvious to read as possible.
The author could have brought the examples forward, instead of just providing links to urllib3 and the Set type.
You can iteratively change a smaller, terser code, must faster. There's a reason prototypes are often written in languages like Python.
In C++, would you write the same structure using a single line for-loop and ternary operators?
The first is that somebody else will eventually have to work with it, decipher what the heck I was trying to achieve, and generally get tripped up and slowed down by it.
The second is that eventually I may have to work with it again, and sit there scratching my head and wondering what I'd been smoking when I wrote instead of focussing on the changes I need to make.
The point is to code for maximum clarity to humans whilst still achieving the desired functional and non-functional outcomes on the machine. You will use varying terseness as part of doing this because terseness isn't the point: it's just one of many means to your end.
It's not "cleverness", but rather the ability to express the intent as succinctly and clearly as possible. Higher level languages allow you to do that, lower level languages don't. Obviously, Go is not a very high level language because it forces you to write explicit loops instead of using comprehensions or functional algorithms. Loop is a primitive concept. All other factors being equal (i.e. performance) you should always prefer comprehensions or algorithms over explicit loops.
for(i=0,total=0;i<resultlen;i++) {
int price = results[i].price;
total += results[i].selected && (price&1) ? price : 0;
}
Objective thinkers tend to go looking for existing, established boilerplate or frameworks because of the immediate leverage that is provided. They may think of too much subjective thought or originality as ridiculous NIH or wheel-reinvention. Their struggle is in the weighty task of grokking and maintaining third-party resources.
Scott Adams has kind of unwittingly uncovered this dichotomy with his own inventive, subjective thinking. He considers more objective thinkers as "word thinkers" or some name like that, because they go looking up the established names of things rather than understanding the things themselves.
Also Chris Langan similarly trashed Jeopardy contestants in a news interview, because they aren't subjective thinkers like he is; their minds are just filled with objective facts about various things. Well, true for him; his bias dictated the way he developed his own unique universal model--he prefers and has a gift for subjective thought.
Really though, we need both kinds of approaches.
And words are amazing. All these complex and nuanced concepts we're talking about can be expressed and shared almost instantly because of the words we share. If we had to invent words as we communicated, or worse yet, had no words, there would be no discourse because the sun would go down before we got anywhere.
It is true that words are set in their definitions and intuitions, much in the way that scientific models are and common sense is. These can be considered objective facts.
But challenging these facts and inventing new words is just an important part of the process. Some have more success with it than others, but not looking up what is already known is just wasting your time. Same with not using words that already mean what you mean.
Subjective thought leads to objective facts with fact checking. Objective facts are then better shared, and better learned, because learning is far easier than arriving at them yourself. Learning just takes communication, which is proportional to the complexity of the information. Inventing or discovering new facts requires the entire lengthy and laborious process which is will inevitably include experiments and fact checking and evidence validation if we are to be scientific. If the problem can be solved with pure thought, chances are it was a philosophical one, although philosophy too highly relies on existing language.
The thing with code is that you can take those existing patterns and make them a first-class concept in your program. You're still in "objectively thinking" - still reusing existing code - just in a more compact and safe way.
If having code be proportional to complexity was a good idea, we'd all be writing in assembly and subroutines would be frowned on.
Boilerplate is when you requires more code than the complexity requires, usually because the language doesn't provide a sufficient opportunity for abstraction.
Note that, if we define a '->' operator that roughly expands to the following:
(x, err) -> f :
if (err != nil) {
return nil, err
}
return f(x);
DoSomething -> SomethingElse -> MoreWork;
This is the design philosophy of Go, as far as I can tell.
process.limits.soft.cpu = "4m" -- limit to 4 minutes of CPU time
process.limits.soft.core = 0 -- and no core file
> I’d take any opportunity to write terse code. It feels good to be clever
> I noticed that when I wrote Python code, I would often go out of my way to avoid messing up a beautiful terse set of lines by moving logic to places other than where it might naturally belong.
I can see how a less expressive language might help with this problem, but I think a better solution would be to rein in the cleverness and focus on readability rather than shortness.
In unison they sort of tells us why pragmatism is often good when we start writing code, but how we must always try to push towards finding (good) abstractions as the work progress.
That's not what boilerplate code is, and DRY is not the way you get rid of boilerplate (the article claims it is a part of the solution).
boilerplate is a repeated preamble that's necessary and unchanging in order to get down to the business of actual logic.
An easy example of boilerplate would be include guards in a C/C++ header file, or 'use strict' in perl.
For example, the author says this:
> When I write Go, I noticed that my tweaks almost always go exactly where they naturally belong. My boilerplate doesn’t stay untouched!
If you're "touching your boilerplate" while writing code that affects the logic of your system, then it isn't boilerplate.
Those examples are just clear indications that languages sometimes have really terrible default behaviors. Namely, Perl should have been strict by default (with a "no strict" if this didn’t perform well enough), and C++ should have at least added #import.
Go has so many truly bad design decisions (lack of generics being the most damning) and it's sad to see its fanboys try to argue that they're actually good.
I was expecting a discussion about boilerplate but the author writes about how ugly/bad is code that does a lot of operations in one line.
Then he introduces the concept of "proportionality", and according to it, assembler must be the best language since the program statements are directly proportional to the operations to be performed...
Suddenly high level programming is no longer good stuff...
Meanwhile no real discussion about boilerplate!!
Whenever your code shows implicit patterns that get repeated but cannot be abstracted away using code, you have boilerplate and boilerplate is almost always a bad thing. Your choices are then:
1. Pretend those patterns as a good thing.
2. Switch to a more powerful language and destroy the boilerplate either because it isn't necessary anymore, or because you can abstract it away with, for example, a macro.
> Then he introduces the concept of "proportionality", and
> according to it, assembler must be the best language since
> the program statements are directly proportional to the
> operations to be performed...
>
> Suddenly high level programming is no longer good stuff...
Function calls are exempt from that, because everyone understands that a function can do arbitrary amount of work, and one needs to know what it does (by reading its name and documentation). The cost is considered to be a "function call".
Yes, Assembly also has high proportionality, which is a benefit to it. But it doesn't mean all other languages are bad. These are tradeoffs.
The point is that it's a beneficial property of a language when you have code length that better corresponds to the amount of work it's doing.
Why they should? It has no reason to be that way.
If anything, 1 line should do 1000, if possible. Some lines will do 1, others 10, others 5, others 100. "Proportionality" adds no value and, quite the opposite, discourages programming at the highest level possible.
I think it helps improve readability of the code.
Good boilerplate is boilerplate that usually gets changed over time (and in different ways each time) as we iterate over the code.
It took me a while to come to this realization (and I explain the reasoning and consequences in the post), so I hope it's helpful to other people who might not have found this intuitively obvious.
Also perhaps "terse" or "clever" are not the best words to describe the counterpoint—I welcome edit suggestions.
With allegedly idiomatic Go:
b := 0
for _, r := range results {
j, k := r[0], r[1]
if !k || j % 2 != 0 {
continue
}
b += j
}
// I see ... given an iterable of results pairs, we just sum the
// first elements that are even for the pairs in which
// the second element is truthy!
Python:
b = 0
for j, k in results:
if (not k) or (j % 2 != 0):
continue
b += j
# I see ... given an iterable of results pairs, we just sum the
# first elements that are even for the pairs in which
# the second element is truthy!
b = sum( ( j for (j, k) # ... sum the first elements
in results # ... of results pairs
if (k and j % 2 == 0) # ... if first elt is even and second elt is truthy
) )
It's further argued that lowest-level-constructs-available scaffolding that may occasionally need to be added in the moment while tinkering is worth keeping around forever despite any tax it imposes on the creation, the reading, and the ease of modification of the codebase and a great advantage of Blub (err... Go) is that it forces this to happen by only having such constructs in the first place.
* What are b, j, and k?
* What's so significant about even numbers - or is 2 a magic number that could just as easily have been 3?
* What can I assume about results? Can it be empty, null, 10 elements, 10 million elements?
With proper naming and a good, brief chunk-level comment, I could tell at a glance whether this chunk is likely to be hosting my bug or not. 5 lines vs. 8 lines don't play much into it.
Actually, this really isn't what I think of when I hear "boilerplate". I think of, for example, the adapter pattern, or having to build up a finicky set of states every time some common task is done.
> With Python, you can write the exact same code that asks for the same mental gymnastics.
The second Python snippet would idiomatically be written on one line and that's what the article is railing against for unclear reasons.
The comments would not appear in source code, they're supposed to illustrate the thought process of the reader as they encounter the source code.
I find myself refactoring Go code a lot in patterns just like you describe (internal functions for error handling), but for me it feels like a throwback to my imperative programming era of the 90s, using Pascal and C. I'm still old enough that it automatically clicks in my mind as clean code, but I also know it's hopelessly verbose.
Error handling doesn't have to be this verbose. In other static languages that eschew exceptions you could do something like this to achieve the same result:
let result = try!(DoSomething());
Now, can you refactor it without throwing away all that nice terseness? Sure. Let's say we need to wrap the error value:
let result = try!(DoSomething().map_err(|e| MyError.wrap(e)));
Go's verbosity is even more cringing with the loops, and this is actually a place where I rarely need any refactoring in my experience. Is your experience any different? Can you give an example of you ended up refactoring loops?
As evidenced by none of the discussion hitherto noticing that your Python and Go examples actually perform a slightly different process? Was that deliberate? ;-)
(I rarely use Python and never Go; yet I noticed quite easily that the Python code is summing the selected odd elements, while the Go code is summing the selected even elements.)
Good catch. I like abstract, declarative and structured code because if nothing else, it requires some level of engagement to write which has a side effect of reducing silly mistakes.
Edit: I changed it.
If you're only using the boilerplate for a new project every few weeks, it's probably overkill to set up some abstraction with variables and templating language.
Granted, I'm using "boilerplate" here in a more general sense than you did in your blog post.
It gave me some thought, but ultimately I think don't think it is a correct argument.
Ain't nobody (= most people) got time for Hegel :)
I suspect that's why some of the very "expressive", i.e. dense languages, aren't widely adopted.
- You clearly state the atomicity, that this thing happening in 1 long line is specific thought, intent, desire - it shouldn't really be on multiple lines - its easier to remove/comment out intent (as it is always self contained), replace it, optionalize it and reuse it.
- Scrolling the code is damaging to the focus.
- Once you learned the specific methods used to create terse code seeing it repeated all over again you will become more proficient in reading the code all around.
- Its way more fun. Its like poetry in programming - why would we all write dummy boring stories. Language is everything (obligatory babel 17 mention: https://en.wikipedia.org/wiki/Babel-17)
This point makes sense, however I think I like to have my cake and eat it too. Java 8's stream constructions can be relatively terse, but if you need to switch back to a for loop, IntelliJ has nifty right-click menu items for doing so. I think the argument made by TFA is at least related to John Carmack's take on inlining vs. abstraction: http://number-none.com/blow/john_carmack_on_inlined_code.htm...
It's very unfashionable advice in some circles, though. It sounds just enough like "go back to writing unstructured copy-paste spaghetti" to trigger some people's kneejerk. But the ground rules of "don't take subroutines, don't jump backwards" actually keep it from getting too far out of control, because they align you towards code that only uses enough flexibility for the task at hand, hence there's a lower likelihood of missing something. Very good practice for greenfield code where you might reflexively add a generalization that you don't use. I can still get better at it.
Edit: And one of the most important things it taught me, is that I can use scope blocks and a comment instead of a function, and it'll be sufficient.
> Maybe this is subjective but I posit that the code’s shape looks more like what it is.
Indentation isn't merely stylistic or syntactic, but it tells us useful things about code at these lower levels. A more complex code shape should convey more complex logic, and we should want that complexity apparent to us...at least sometimes. It looks great to have a nice, terse one-liner, but visually flattening possible control flow paths feels really nasty.
That being said, the most tiresome code is this kind of step-by-step error checking.
if (! is_numeric($id)) {
exit('The ID is not a number.');
}
$db = open_database();
if (! $db) {
exit('There was an error connecting to the database.');
}
$query = db_query($db, 'select * from widgets where id = $1', $id);
if (! $query) {
$db_error = get_db_error_msg($db);
exit('There was something wrong with the query: ' . $db_error);
}
$db = open_database();
$query = db_query($db, 'select * from widgets where id = $1', $id);
Don't get me wrong. I write JavaScript to constrain user input, with friendly directions and error messages. And I set tight types in the database and add database constraints where I can. But I don't write all this step-by-step checking in the middle layer (PHP, Python, Go, what have you). So if the user somehow bypasses the JavaScript and tries to insert text into an integer column, the database will simply refuse, and the resulting web page will be a very ugly error message, which is what they deserve.
If the database is down, it's going to be ugly, regardless.
I admit that this strategy works okay for non-life-threatening, SQL-backed web apps and such. You may need to do something tighter in your work. But I wonder if there would still be something more elegant than the every-other-line error checking shown in the article.
Yes, it makes you zzzZZZz. If you don't care for errors anyway. However, when developing a large reliable software project you need to think through the possible error cases and handle them appropriately. Even if that just means logging them correctly and exiting so the error can be easily tracked down later (for example: was it expected/unexpected? What where the parts of the context that can't easily be collected by a runtime?).
And in a large project, it's not that bad. How many places are there where you setup the database connection, an OpenGL context, a network connection, etc? Just build the abstractions appropriate for your project once. Simply crashing backwards through the function calls with a generic exception and a stacktrace is not always good enough.
Basically, every module or function or whatever has a client. Sometimes those clients are other parts of your own system, sometimes they're end users outside of your control. Regardless, at significant barriers of abstraction, I try to catch errors, and translate them into new error types/representations that would be relevant or actionable by the intended client of that code, with the original error provided as a member. If you follow this pattern, the logic you write is free of error handling code, error handling code is centralized at abstraction barriers, and at any point you should have few errors that you wouldn't know how to respond to and probably just give up and bubble to the top.
Factor out the error checking into a function:
onerrexit(is_numeric($id), 'The ID is not a number.');
onerrexit($db = open_database(), 'There was an error connecting to the database.');
...
So yes, outside of pure functions you will always need special-snowflake error handling.
I can't see how anyone could prefer the Go version.
And I don't know if python allows this (it may not due to the significance of whitespace), but your proportionality problem can be easily resolved by placing each different clause on a different line, like this:
sum((j if j % 2 else 0)
for (j, k)
in results
if k)This type conversion madness needs to stop. Integer 1 and Boolean true mean different things.
It's one of those rare features that I wish other languages would adopt from Java. Possibly the only one. :/
especially in the case of modulo, it often makes sense to just treat the result like a boolean, because it practically says "divisible yes/no".
theres nothing i remotely like about java and its great that python isnt trying to be like java, but those are just opinions.
is your complaint about the implicit typecast or about the fact that integer can be cast to boolean at all? if its the former then python is just less pointlessly verbose than java. the latter may make sense to complain about.
if you want to sum over even numbers, just do:
sum([j*2 for j in range(0, end_range/2)])
and yes, brackets supercede pythons "whitespace is syntax". if you need linebreaks where, ordinarily, linebreaks are not allowed, just write brackets.
Given an iterable of pairs, sum the first elements of the pairs for which the first element of the pair is odd and the second element is truthy.
What the code you're suggesting does is, to be polite, unrelated.
i agree that i misinterpreted the codes intention.
Programming is more than that.
This is best: sum(range(0, end_range, 2))
;)
can they be great artisans? sure.
why do i think its not better? i said im not sure. i think the two options are about equal in readability, up to taste. yours is probably more efficient in terms of cpu cycles.
whats your sticking point?
It's not right. It doesn't give you the value 10, as you'd "expect".
2
2+4 = 6
2+4+6 = 12
if you sum even numbers, starting from 0, "10" is not a possible result. i dont know what youre trying to do there.
So it skips 10...
if you do sum(range(0,10,2)), 10 isn't reached either.
why are we having this discussion?
[j2 for j in range(0, (11+1)/2)]
Your code is wrong.
sum(range(0,11,2)) works as expected, [j*2 for j in range(0, (11)/2)] does not.
You seem like you have issues.
is what i'd call it.
If you abstract things but don't give meaningful enough names, you loose everything (because your callers are still forced to lookup the definitions of your abstractions, and now this definition can be hard to search for).
b = sum((j if j % 2 else 0) for (j, k) in results if k)
sum [j | (j, True) <- results, odd j]
filter snd >>> map fst >>> filter odd >>> sum
The Opportunity for correction is an interesting idea but the authors issue mostly come from a lack of referential transparency. In a pure language refactoring wouldn't be an issue so one could argue this is a symptom of too little abstraction, not too much.
Slightly less wise-ass, I couldn't come up with a use case for this code that'd give sane variable names. Third attempt:
sum $ do
(element, True) <- results
guard (odd element)
return element
In regard to boilerplate the author brings up an interesting point that with a good language, the boilerplate can effectively disappear once all the code has been written. If people are worried that boilerplate violates DRY, I think they worry too much.
The best way I've found to determine whether a codebase is good or not is to take someone who's never worked with the code before and ask them to fix a bug or make an enhancement. If that's an excruciating task but would have been easy in a greenfield project, the code sucks.
It seems we have scientific data showing that the number of code lines one can review per day doesn't depend on the programming language. This means one would spend a lot more time reviewing this C++ code if you looking at the manually-rewritten-to-assembly version instead.
This is a very good argument supporting programming at the highest possible level of abstraction, which is the exact opposite of boilerplate and verbosity.
So please let's not rationalize/stockholm-syndrome over the limitations of any programming language ; boilerplace and verbosity are always bad.
is_even = lambda n: n % 2 == 0
score_result = lambda j, k: j
is_interesting_result = lambda j, k: is_even(j) and k
a = some_variable + 42
b = sum(map(score_result, filter(is_interesting_result, results)))
b = sum(j for (j, k) in results if is_even(j) and k)
Using functional programming paradigms like this often advised against in Python.
[1] https://www.python.org/dev/peps/pep-0008/#programming-recomm... (search for "Always use a def statement")
I'm not really concerned with PEP. I'm concerned with if the idea being expressed is cleaner then the original. My first iteration was cleaner then the original comprehension (despite the comprehension being PEP8). It also let me see what I was really doing and I was able to turn my map and filter into a much nicer list comprehension.
Also, pointing out lambdas instead of defs is silly. It was done for an example. Most of the time you'll be operating on data structures with built in functions or supporting functions you will have written making it cleaner to use the map and filter paradigms.
money = sum(user.get_payed_balance() for user in users if
user.has_paid_bills() and
user.is_still_subscribed())
clients = sum(map(User.get_payed_balance,
filter(User.is_still_subscribed,
filter(User.has_paied_bills, users))))
> But perhaps even more important, it doesn't even work since a map-function can only take one argument (not two).
The concept still stands. Just use `from itertools import starmap`. Starmap is like map that calls the * operator on arguments.
>>> a = lambda a, b: 10
>>> list(map(a, ((1, 2), (3, 4))))
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
TypeError: <lambda>() missing 1 required positional argument: 'b'
>>> list(starmap(a, ((1, 2), (3, 4))))
[10, 10]
I'm not concerned with if it's "Advised against". I'm concerned with a) will someone be able to understand this better then the implementation that currently exists) and b) will I be able to come back in 10 years and tell what this is doing if need be.
You should check out this talk [1] before going around trying to use PEP8's inconsequential clauses to call something "horrible" code.
So for me it is not always bad, just usually, with some specific exemptions that I think are done well.
I call BS. Consider these 2 Go lines:
heck_if_I_know_what_this_does()
heck_if_I_know_what_this_does_either()
>Is this good? I believe it is. I believe this code is more proportional to the complexity of the work that it’s doing.
We invented programming languages on top of machine code so that our code can be LESS proportional to the "complexity of the work that it’s doing".
>When reading proportional code, it’s easier to notice where the interesting bits are.
Not really. It's the inverse: the terse code only keeps the interesting bits. The "proportional code" gives you all the non-interesting, and getting in the way, mechanics.
>When a friend was reading through the source code for ssh-chat, he was surprised that I implemented my own Set type. I explained that while Go doesn’t have a built-in Set, it’s just a few lines of boilerplate to make your own on top of a map. In fact, I noticed that my version of Set evolved to be fairly specific to how it was being used. In retrospect, I’m glad that I was tweaking my own implementation iteratively rather than spending time working around whichever limitations a generic library might have had.
Pray tell, what "limitations" a generic Set would have?
And why does those kinds of post remind one of Stockholm Syndrome?
True, but that's not what the article's arguing.
> We invented programming languages on top of machine code so that our code can be LESS proportional to the "complexity of the work that it’s doing".
No, we invented programming languages so that our code can be smaller and less arcane, not less proportional. The actual point of the article, which I believe you missed, is that if we focus on 'cleverly' compressing large amounts of logic into dense nuggets of code, it interferes with our ability to appropriately factor the code so that it produces the desired behaviour in the simplest and most readable way.
The article also states (and I agree) that if your code is wildly variable in the amount of semantic content per line of text, it's much harder to properly grok and (more importantly) to modify what's going on.
Reading it twice, I think this is a very veiled argument against generics in Go on the grounds that well, uh... proportionality?
I agree with the post above you. I do not like the premise of this or of Go. I see it as a deliberate attempt to walk away from empowering programmers. It's not solid technology, it's the technical manifestation of business decisions.
Smaller also means less proportional to the work that it's doing, since the abstraction, sugar and techniques we use to make code smaller do not yield the same compression ratio from the underlying instructions in all cases.
So, smaller = less proportional -- unless you suggest we should also strive to artificially constrain how succinct we make our code so that there's a constant ratio everywhere.
That is, so that 10 lines of code in a language always are (or at least, always strive to be) 10×n times of machine code. I don't see how that's feasible -- some abstraction could yield a 1/100 less code than machine code in one line, while another can yield a mere 1/5 or 1/10 less code.
But even if we could achieve that, I don't see it being even useful. Having higher level code strive to be more proportional to machine instructions is not necessarily telling us much about the cost of such code. It might tell us that it's complex and slower than another, but that doesn't mean much. It only matters what role it plays in the overall runtime of our program. Just because something A 5 times bigger than something else B, doesn't mean A needs more optimizing. A could be just fine as it is, e.g. if A is called rarely, and B is in some tight loop.
Gauging code by "code proportionality" (the only semi-legitimate use I could think of such a feature) is bogus. We should always profile.
>The article also states (and I agree) that if your code is wildly variable in the amount of semantic content per line of text, it's much harder to properly grok and (more importantly) to modify what's going on.
On the contrary. The part of "What's going on" that matters when it comes for modification is the high level stuff that goes on -- which a more terse code shows better, as it doesn't hide it with boilerplate.
In the end, that point seems both trivial and non-sensical: yes, boilerplate is easier to modify. That's because it doesn't matter much -- it's the stuff that you want to do that's hidden inside the boilerplate that matters.
e.g. I can write a loop to filter 100 numbers in tons of ways. And I can always modify it from one to one of the other ways.
But something like:
value = filter(my_array, (val) => val > 5)
I'd argue the contrary though.
That the "semantic complexity" is less when you only keep intent-code, than when you spread it over 20 lines of boilerplate which you have to visually parse to get to the point.
Your examples are not analogous, because they do different things.
Let's make them implement the same example, Go/C/etc style and "terse" style:
array_2 := make(int64, 0)
over_five_index = 0
for i, val := range(my_array) {
if val >= 5 {
over_five_index += 1
if over_five_index % 2 == 0 {
append(array_2, val)
}
}
}
array_2 = my_array
.filter((val) => val >= 5)
.filter((_, i) => i % 2 == 0)
As for the first example, it would be:
a = b || c
Indeed the author is suffering from it.