LoC is a dumb metric for functions | Hacker News Reader

LoC is a dumb metric for functions

theaxolot.wordpress.com

61 pointsAxol62 comments

LOC is often a rough approximation for complexity. We once had an intern who made some useful things, but he didn’t know how to break anything down. One of them was a 1,000 line perl all as one function. I asked if it could be broken down into something more maintainable and he said no. There were several projects like this.

Knowing At a high level what needed to happen to accomplish what the code did, I know for a fact it could have and should have been broken down more. However, because of the complexity of what he wrote, no one had the patience to go through it. When he left, the code got thrown away.

While 10 LOC vs 50 LOC doesn’t matter, when commas enter the number, it’s a safe bet that things have gone off the rails.

Seems like the article missed an opportunity to talk about testing and MC/DC coverage. I don’t care if the PR is long or short, just show me that you have meaningfully tested how each branch can be reached with the full range of values and behaves correctly. Unit testing is easier with well chosen interfaces and worse without them, regardless of LOC.

I don’t understand either. The complexity will be there no matter whether you split up large functions to smaller ones or not. If you need to understand the given function, you need to read and parse the same amount of code. If you blindly trust naming of functions, you will have a bad time. You need to write smaller functions for other reasons, not because of complexity:

- Testing

- Reusability

- Configurability

- Garbage collector/memory management

- etc

I never understood these kind of restrictions in code analysis tools. These kind of restrictions don’t help overall complexity at all, and sometimes they even make things more complex at the end.

>When he left, the code got thrown away.

Why even merging his code at first place. He was intern so i assume whatever he was working on was not mission critical.

They were one-off projects where he was the sole dev, not part of a larger project. While not mission critical, the 1k line Perl script was helpful at the time. When we ran into a specific issue it allowed for recovery in minutes vs hours. We eventually added safeguards which prevented the issue from happening in the first place.

> While 10 LOC vs 50 LOC doesn’t matter, when commas enter the number, it’s a safe bet that things have gone off the rails.

There are times when even a 1,000 LOC function is the better solution. The best example I can think of involve a switch statement with a very high branch factor (e.g., 100), where the individual cases are too simple to break out into separate functions. Something like the core loop of an interpreter will easily become a several-thousand line function.

Cyclomatic complexity is not equal to LOC and cyclomatic complexity of a switch can be seen as adding just 1 to its enclosing function. Either way, LOC is still not a great metric while cyclomatic complexity approaches a better one.

In my experience, there are very few places where something can't be broken up to reduce cognitive overhead or maintainability. When a problem is solved in a way as to make it difficult to do then maybe it's better to approach the problem differently.

Higher cyclomatic complexity requires more lines (barring bad code style like writing an entire program on a single line).

The inverse is not always true, but often is.

> Higher cyclomatic complexity requires more lines

Often true; That's why cognitive complexity wins over cyclomatic complexity.

eg: Embedded logic doesn't add linearly.

fn() { // 1

  if(...) { // +1
    if(...) { // +2
      if(...) { // +3
         ...
      }
    }
  }

} // cognitive complexity of 7 vs cyclomatic complexity of 4

It's been a while since I've implemented anything around this and was remembering cognitive complexity while writing cyclomatic complexity in the previous response. They both have their place but limiting cognitive complexity is vastly more helpful, IMHO. eg: The above code might be better written as guard-clauses to reduce cognitive complexity to 4.

Interestingly, I read once that Hacker News moderators use a similar metric regarding the level of indentation in a thread. It’s a blunt tool to spot flame wars. More replying = more indentation = more heated = less interesting.

That's a special case

I would argue an interpreter that needs 1,000 lines for its core loop is probably a complex piece of software, comparable to other 1000-line projects I've made.

John Carmack wrote about inlining code: http://number-none.com/blow/john_carmack_on_inlined_code.htm...

> I asked if it could be broken down into something more maintainable and he said no.

This is something AI is good at. It could have shown the intern that it is indeed possible to break such function without wasting an hour of a seniors time.

> LOC is often a rough approximation for complexity.

I would argue that the word you are looking for is "containment".

Is there any real difference between calling a function and creating a "const varname = { -> insert lots of computation <-; break retval;}" inline?

If you never do that computation a second time anywhere else, I would argue that a new function is worse because you can't just scan in it quickly top to bottom. It also ossifies the interface if you have the function as you have to rearrange the function signature to pass in/out new things. Premature optimization ... mumble ... mumble.

Even Carmack mentioned this almost 20 years ago: http://number-none.com/blow/john_carmack_on_inlined_code.htm...

Given that we've shrugged off "object-itis" which really demands small functions and given modern IDEs and languages, that kind of inline-style seems a lot more useful than it used to be.

> If you never do that computation a second time anywhere else, I would argue that a new function is worse because you can't just scan in it quickly top to bottom.

I feel exactly the opposite. By moving "self contained" code out of a method, and replacing it with a call, you make it easier to see what the calling location is doing. Ie, you can have a method that says very clearly and concisely what it does... vs making the reader scan through dozens or hundreds of lines of code to understand it all.

Chapter 12 of Kent Beck's book Tidy First? is about exactly this.

Do not cite Kent Beck as a software authority. Consider this your semi-regular reminder that Kent Beck is part of the posse (along with Ron Jeffries and Martin Fowler) responsible for the software disaster that was the "Chrysler Comprehensive Compensation System". They were also proponents of Smalltalk--which leads to object-itis and the concomitant tiny functions.

As for your "function documentation", the primary difference between that assignment and a function is solely ossified arguments. And ossified function arguments is something that I would put under "premature abstraction".

At some point you can pull that out into a separate function. But you should have something that causes you to do that rather than it being a default.

> Do not cite Kent Beck as a software authority.

I didn't. Rather, I stated my opinion on the subject and then noted that there is a book that discusses the same (so clearly I'm not the only one that thinks this).

> As for your "function documentation", the primary difference between that assignment and a function is solely ossified arguments.

No, it's not. It's that the code being moved to a function allows the original location to be cleaner and more obvious.

I find out useful to split things up sometimes to reduce the number of things in scope at a given time. I've found that this both makes it easier for me to reason about, and anecdotally over the years it seems to correlate with my coworkers finding it easier to review as well, although I suppose it's possible that there's a bit of selection bias in terms of teams being more likely to hire people (and individuals to join teams) with similar mindsets.

> If you never do that computation a second time anywhere else, I would argue that a new function is worse because you can't just scan in it quickly top to bottom.

I do the computation in my head each time I read it. If it is a function I can cache its behavior more easily compared to a well defined block of code. Even if it's never used anywhere else, it's read multiple time, by multiple people. Obviously, it has to make sense from a domain logic perspective, and not be done for the sake of reducing lines of code, but I have yet to see a function/module/file that is justified in being gigantic when approaching it with domain logic reasoning instead of code execution reasoning.

An example I'm dealing with right now...

At my job, we're not a SaaS. Our software is distributed as cloud images that customers run in their own environments. I need to run regular vulnerability scans on built images.

So, I need to do several things:

- Stand up EC2 instances with our image

- Run our software setup

- Start the scans

- Wait for the scans to complete

- Destroy all the instances

- Download the scan reports

- Upload the reports to S3

- Create Jira tickets based on findings

Now, I COULD easily write it all as one big function. But I tend to prefer to have one "main" function that calls each step as its own function. When looking at the main function, I can much more easily tell what the program is going to do and in what order.

In other words, I much prefer style A from Carmack's post.

After having this discussion in so many professional settings, I'm starting to think that this is just something that divides people. Maybe our brains are wired differently. What's important for me is that the call graph is simple. Second most important is that data structures are easy, or at least fits the problem. The amount of lines in a function is a distant third.

It could have been me behind one of those thousand line functions. Some tings lends itself naturally to this style, such as interpreter-like loops and imperative business logic. There's always that someone that argues that splitting up the code with function calls would somehow make it more "maintainable".

I'm reluctant to do it, because the code is identical, only non-linear which means harder to understand. Combine this with global symbols and it's a recipe for accidental complexity. This is so obvious that I for a long time thought people with the opposite opinion were just cargo-culting some OOP "best practice" mess which has doomed so many projects, but maybe they're just wired differently.

I guess I don't really see a reason to have a 1000 line of code function that's not broken down.

Although I personally despise the massive call graphs, my rule of thumb tends to be if I spend over 10 minutes trying to hold all the logic in my head, something is wrong.

Especially with imperative business logic - surely it makes so much more sense to break it down into functions. If you're on call at 2am trying to solve a function with imeprative business logic that someone else has written and has over 1000 lines, you're going to hate yourself, and your coworker.

I think you're right. I'm on the opposite side of it, but I have a colleague who routinely writes functions that are hundreds of lines long, with 30-40 local variables inside each function. I've come to realize that he does so because his brain allows him to do it. He has a brain that is more detail-oriented than my brain. Where I naturally see a problem as a "tree" of sub-problems with their own details, he naturally sees one large problem with 100 details. He naturally writes code to accommodate his own cognitive style, while I write code to accommodate my own cognitive style.

Moreover, I've come to realize that my colleague is not adverse to using abstractions if they are well established, and if they already exist. But he is (much) less inclined to invent or write new abstractions than I am. As you have concluded, I have also concluded that this is actually a matter of cognitive style, more than it's a symptom of "slop" or "cargo-culting of best practices".

I had a colleague who wrote long functions like that. Over time I came to think of him as an inconsiderate asshole. Unfortunately he was me a few years earlier. I gradually learned the value of organizing code in short sweet methods. When that style become old code, I and my other colleagues came to like that crusty old coder better. I also thought it was a matter of cognitive style but for me it turned out to be inexperience.

Do you write comments in your code? If you ever have a long function that like

    ....
    // Here's where we apply coupon codes 
    ...

It makes sense to me to break that out into it's own function, preferably one with no side effects that I can write a test for if I need to. I can give it a name related to the business logic.

Inevitably there will come a new special case where, for example, where Gold Members will get to use double coupon codes on Thursday. You'll start by looking in `checkout()` where you'll see either 1000 lines of other special cases, or maybe 30 lines with function calls like `foo = applyCouponCode(shoppingCart, couponCode)`.

For me, it's easier to see where to start with the latter.

Comments are perfectly fine on their own. It's common that a block of code can be given a higher level explanation in a comment, but wouldn't be a good standalone function because it depends on the context too much and so the function has absolutely no chance of ever being used a second time.

It's okay if a function is only used once. It's worth it to separate it from the context it doesn't need.

Well, twice, including the test. If a block is complex enough to warrant a high level explanation, you might as well capture that intent in a test too.

Edit: in my example, `applyCouponCode` takes in a `shoppingCart` and a `couponCode`, so you know that's all you need to apply a coupon code. You'd change it to something like `applyCouponCode(shoppingCart, couponCode, user.memberStatus)` so you can tell at a glance that `memberStatus` has something to do with how coupons work. You wouldn't want to pass in the whole `user`, because giving the function more than it needs makes it hard to infer what context the function needs and, therefore, what it does.

Tests are code. If the function is easy to test then it's more likely to really be worth having.

I think a better measure than simple LoCs is to cluster the local variables by use. If you get entirely separate clusters or large clusters that only share a few variables, then you should break them up at the cluster boundaries.

Exactly, at that level someone is not thinking about how someone else will read and understand the code.

> Knowing At a high level what needed to happen to accomplish what the code did, I know for a fact it could have and should have been broken down more. However, because of the complexity of what he wrote, no one had the patience to go through it. When he left, the code got thrown away.

And today an LLM could probably have refactored it fairly well, automatically.

I worked closely with an engieer that would write these massive long functions; but they read clear and linear. They were really easy to follow and understand.

He'd of course develop functions and macros to avoid repeating themselves.

It just worked for his style and problem solving approach. He knew when to stretch a function. So, I'm convinced long functions can be done elegantly, it's probably uncommon.

> LOC is often a rough approximation for complexity.

Setting aside the anecdote, this statement is key. If you exclude deliberate efforts to game the LOC metric, LOC is closely correlated to complexity, however else you measure it (branch count, instruction count, variable count, rate of bugs, security vulnerabilities, etc). Unlike any other metrics however, LOC is dead simple to measure. Thus, it's an extremely useful metric, so long as you don't set up an incentive for devs to game it.

Worth noting that "cognitive complexity" may mean SonarQube metric – a metric that is not widely recognized by industry, created by SonarQube employee as (imho failed) attempt to address "issues" with "cyclomatic complexity" (principled, intdustry recognised metric). It'll count things like nullish coalescing and optional chaning as +1 on your complexity which makes it unusable with jsx or ts code.

With low thresholds ("clean code" like low) both LoC and "cognitive complexity" (as in SQ) are bad measures that lit up in red large percentage of otherwise correct code in complex projects. The way people usually "solve" them is through naive copy pasting which doesn't reduce any cognitive load - it just scatters complexity around making it harder to reason about and modify in the future.

>Use your judgment, and don’t be bullied by people who prescribe specific line numbers.

Ironic way to end an article that repeatedly belittles the target audience.

Use your judgement, don’t be bullied by people who prescribe using your judgment.

It's not the LoC I care about, it's the logical separation of concerns and testability. Large functions usually do many things which makes them really hard to test. Also, just mashing all the things in a single function is indicative of the author not having a clear picture of what problem he or she is dealing with.

I have been programming for 30 years and, while I don't consider myself a great programmer, I don't like large functions. In my experience they usually fail at clearly expressing intent and therefore make the code a lot harder to reason about. There are, of course, exceptions to this.

It may be a bad metric, if you're being proscriptive, but it's a great heuristic. If I see a 500-line function where I'm not expecting one, I'm going to pay a lot more attention to it in the PR to try to figure out just why it isn't shorter.

> Locality is the design principle of grouping related things together. It’s a mental shortcut or heuristic that allows the mind to assume a higher degree of independence from the code block being read, easing the mental burden. Spreading code across your codebase diminishes locality.

> Linearity is the idea of having instructions be read one after the other, from top to bottom, just like how code is executed. Having to jump back and forth from function call to function call reduces linearity.

1000 times yes.

It's about structures. A long function is unfortunately "flat" at the first glance. Even if there are inherent structures, it usually burns a lot of brain power for a human to abstract structures out of "flat". Being flat is a more acute issue for LLM to understand. I think in the LLM era it's more important than ever to keep things short and structured.

For a longer, more detailed take on this see:

https://www.youtube.com/watch?v=bmSAYlu0NcY

which is for the book:

https://www.goodreads.com/book/show/39996759-a-philosophy-of...

This needs more than just an upvote — that's one of the best software engineering books I've ever read!

The whole point of a rule-of-thumb is to be dumb. There are very few good reasons for a 500 loc func, and the code reviewer won't want to verify it's as simple as you claim. You could still have a 50 loc func that's overly complex, and they'll complain about complexity instead of length.

One consideration is that about 40 loc will comfortably fit in one code split, and it's nice if one func will fit in 1-2 splits. Past that length, no matter how simple it is, the reader will have to start jumping around for declarations. That might be ok at first, but at some point the ratio of offscreen to onscreen code gets too big.

I wonder if there is a way to see how physical monitor quality and size improvements have led to more complicated code, nevermind moving off of punch cards.

The type of monitor doesn't matter a whole lot because it's really limited by human eyesight. High-res monitor will enable rendering things tiny if you disable hi-dpi, but then it's unreadable. If you use a big 8K TV to display everything larger, you have to sit further away to comfortably view it. If you add more monitors, at some point it becomes too hard to look at so many things at once.

Personally, my setup has shown the same number of vim rows/cols on my 4K monitor vs my 1080p one (or whatever you call the 16:10 equivalents).

The effective resolution of the eye should be an angular value. I'd say that the smallest comfortable size of text characters for me is about 1/8° wide. I see the width of my 14" FHD laptop screen at about 30°, so the full width contains about 240 characters. This is a large upgrade from 80 columns of VT220 or VGA, and almost twice as much as the densest VT220 mode of 132 columns.

My 28" 4K monitor is exactly like four 14" FHD screens put together. It offers me 480 columns (a bit fewer, because window borders and scroll bars of multiple windows).

So indeed, better screens allow me to see much more text than poorer screens of old. There is a limit, but definitely it was not reached 30-40 years ago yet.

I can read small things much better on a high res monitor.

Back in the day we were rocking 640x480 on a CRT. 800x600 was what the rich kids had.

The winning feature of LoC as a metric is that it can be "eye-balled" easily without tools. Cyclomatic complexity is a more interesting metric; however, it's not the easy, quick, and rather dirty spitball that Lines of Code is.

A personal guideline for a lot of stuff is that a function may be too long when people add comments to mark what sections of it do. (ofc not really a hard rule). I just think it's easier to see "oh this is calling the load_some_stuff function, which I can easily see returns some data from a file." Rather than <100 lines of stuff, inlined in a big function, that you have to scan through to realize it loads some stuff and/or find the comment saying it loads some stuff>. That is to say, descriptive functions names are easier to read than large chunks of code!

smaller functions are also usually easier to test :shrug:

Sometimes I write long functions because I do not have time to write a shorter function. Eventually, if I have time, I go back and refactor a much shorter and tidier version. While later versions may be somewhat more efficient computationally, it is mostly a no-op beyond the aesthetics.

What is the LoC for that function? The first implementation or the final rewrite? They express the same thing.

Love the article thanks. to me it reads really reasonably. i work in a verbose language (C) and there its too easy to try and optimise things away to 'save typing' to a point it becomes actually more of a burden than an optimization.

Some good advices in here on what balance to strike with some clear examples. Always a good reminder :). thanks for the writeup!

I always find it mildly funny/annoying when reading macros which only appear to save the pointer access caracters

#define G(l) l->_G

What leverage have you achieved there buddy

That can be useful if it is l->foo.bar.baz._G .

What an arrogant, wordy, and partly wrong take on this matter.

No, LoC is not a dumb metric. In the same way that cognitive complexity, code coverage, and similar metrics aren't. CC is definitely not "superior" to LoC, it's just different.

These are all signals that point to a piece of code that might need further attention. It is then up to the developer to decide whether to do something about it, or not.

You know what's dumb? Being dogmatic, one way or the other. There are very few hard rules in software development—they're mostly tradeoffs. So whenever I hear someone claiming that X is always bad or always good, I try to poke holes in their argument, or just steer clear from their opinion.

My feelings on LoC really depend on what you want to measure. LoC tends to correlate with # bugs. A 100-line function is probably going to have more bugs than a 20-line function that does the same thing.

I don’t think functions inherently need to be small though. 5 20-line functions are just as buggy as that 1 100-line function. Bugs scale with LoC, not how well you’ve broken functions apart.

The lesson is really just to avoid overcomplicating things. Use less LoC if you can. You should also avoid overly-shortening or cleverness but it tends to be the lesser evil.

I’d hazard that the 5 20 line functions are sometimes slightly less buggy, since the function signatures act as essentially an enforced comment on how the implementation is done and perhaps which state is accessed/modified. Versus mutable variables and if-statements in the larger function, that is.

I’d say more testable. I haven’t seen any metrics on amount of bugs vs composability of functions though. There are metrics on amount of bugs vs LoC.

I have a personal preference towards the “more functions” approach too but I want to separate out my emotional feeling about what’s better vs what’s actually supported with data.

LOC is an important metric for cost. There's a certain budget of what a person or team can understand/manage.

Citation needed. That LoC is the thing that causes bugs may be a dumb idea, but the best available evidence is that it's true.

Suggesting that it's amateur to split up a function that is "too long than X" is spot on, and missing the point at the same time. Amateur coders will also have to read and try to leverage the code. That's why it has utility.

Reply on news.ycombinator.com