Rust changed me as a developer
blog.startifact.com
blog.startifact.com
It is a scale, and there are other languages that go beyond Rust, with dependent types, effects, capabilities, linear types, formal profs,...
Similarly how folks get impressed with Go compile speeds because they weren't around in the last century to have used Turbo Pascal, Modula-2, Oberon,.... compilers
Or assume most Zig improvements over C are innovations, when even languages that predated C like PL/I had most of the same safety features.
The issue is lack of understanding where we come from.
One could also argue that the Rust ownership system is also unique. It was inspired by other work but the difference between the inspirations and what Rust ended up with is fairly drastic.
Pretty much everything else came fairly directly from elsewhere; ML primarily.
But note that zero-cost async and the ownership system are the worst parts of rust! Ownership and zero-cost async bring enough benefits that people use Rust to get them, but it's always a little bit begrudgingly. There are sharp edges.
OTOH, the parts of Rust that it borrowed from elsewhere are beautiful, IMO. They've used the advantage of hindsight to refine them.
At some point in the future another language will do the same to Rust, refining zero-cost async and ownership. And creating their own sharp edges to add a more novel capability.
As regards ownership, I know of at least one language already that takes the ownership model and streamlines the lifetime annotations out of it. See Loon[1], and possibly March[2].
The issue is that people should understand their history. Instead of saying Rust is amazing because it invented certain concepts, they should be digging in further to understand from where those concepts came. That's not at all a knock on Rust, purely the fanboys who make it their identity.
If you just want generic pure computation: write an interpreter for a toy language. If it’s too large, write a macro expander that can handle C style or TeX style macros.
If you want events and concurrency: write a tmux clone that can work with multiple terminals.
If you want IO but specifically networking: write a toy HTTP server that only supports 1.0.
I am not sure if everyone else looked up techniques online or what for how to do it, or if they skipped that assignment, or made a very simplified solver that only handled some basic techniques. We hadn’t really learned much of anything yet in any of the classes of the course that would be useful for making a general solver. I thought I was supposed to figure it all out on my own, and that it should be able to handle any Sudoku puzzle that a person could solve. And in the end, despite many, many hours of working on it over many days it was still far from a complete solver. And I also didn’t have time to write it in the programming language we were using in the course. I was prototyping in JS and HTML because I wanted to see visually the sudoku puzzle and be able to switch between having the computer find next number and clicking on the number myself, to discover in detail how I was thinking when I was solving Sudoku puzzles, so that in turn I could implement those techniques in code.
I lost a little bit of motivation for the class with that, and dropped out of the class soon after because I found the class to be generally uninspiring and hard to stay motivated for, in spite of being interested in programming in general and having learned programming languages and made little programs of my own already since High School. Later I dropped out of the whole study, as I was struggling to study across most subjects and wasn’t passing enough classes to be able to continue.
My interest in programming continued, however, and I’ve been developing software both professionally and as a hobby with sometimes pretty good results during the years since, and having done so for companies both small and big, and in multiple different programming languages.
But I think the way I handled trying to implement the Sudoku solver is a way of working on things that I already had since before then and that I continued also later, and that that is also part of the reason that things go wrong for me when they do. Both professionally and also elsewhere in life.
The last job I had, I was laid off for not performing well enough. And I think in part I was still working on tasks at that job in some sort of wrong way all these years later, that may be similar in some ways to what I may have done wrong with for example that Sudoku solver assignment. And probably in many other cases too both before and after.
And now that AI is here I’m not really sure how to land a new job as developer at all. Even though I would like to. Combined with the sense that there is too much ups and downs in success rate of what I’ve been working on over the years to be able to present myself in an impressive manner to employers. I value honesty and integrity, and I find it difficult to really imagine that I could sell myself well at this point when I look back at my life so far, in spite of sometimes doing things well, because of the times that projects have not gone well.
If so, that particular shortcoming might be easier to address than many of the other possibilities.
Yes I think that is probably the case, at least some of the times.
For anyone curious, I generate the sudokus by first randomly filling an empty grid in a way that follows the sudoku rules, then emptying random cells where doing so doesn't make the solution ambiguous (a sudoku has exactly one valid solution). The first step to fill the grid works almost exactly like a normal backtracking solver but going through the cells in random order, and for a major speed boost the backtracking is far more aggressive (always resetting half the progress so far instead of just going back one step). Checking if a sudoku has multiple solutions is again normal backtracking but only exiting at the second time it finishes or after going through all combinations. The rest is modern computers being fast and limiting the number of iterations to avoid the super slow worst cases.
A common example I give is a basic chat client/server, to learn how networking code feels, especially nonblocking IO. Especially if you're implementing the same protocol in the different languages since it means you won't necessarily be as biased by what protocol patterns are easy to implement in one language.
Something that has a lifespan longer than a month.
Something developed under time pressure.
Something that is forced to change due to outside real world impacts.
Something that when you come back to it, you find some code that prompts you to say “What the heck was I thinking here?”.
Elixir was somewhere between my 12th-14th language and nothing changed my perspective so much as understanding why the BEAM works the way it does.
(I wanted Prolog to blow my mind too but I could never figure out what the hell was going on!)
My favorite side quest that I definitely recommend. Physical parallelism is so fun to reason about.
I had the advantage of sitting next to a proper professional to bounces questions off of, so I was able to get an SPDIF decoder to work in about a week! It was a very interesting lesson in the costs of abstraction, and why you will never really find abstraction layers in hardware.
From the compiler construction standpoint, it's grossly underspecified, to the point that our customers' designs were way more expressive than the spec.
And from the users' perspective, it's confusing, quite limited, and you end up having to resort to hacks to achieve what should be part of the language... Fortunately I can't remember any particularly bad examples, it's been 15 or so years since I left for greener pastures.
Still, I sometimes wonder, how much of HW design is influenced by the language decisions, these ugly hacks, and the emerging (System)Verilog design patterns? Could it be measurably better in a different world?
I think 25 years ago I would not have believed we would ever fix (La)TeX but here's Typst.
It's going to take a long time though. All IP vendors provide Verilog, often encrypted which only the commercial tools can legally support. And there's a very very long tail of SystemVerilog features that you need in ASICs - coverage, timing, etc.
This is interesting to me, in the very little HDL I’ve done there were abstractions: you set up modules with inputs and outputs that provided a consistent interface despite complexity inside the model. Is this not what you mean by abstraction layer?
The rest of synthesizable Verilog is a pretty thin abstraction over muxes, d flip flops, and adders, and other basic gates.
That being said registers is a pretty big abstraction on top of what actual circuits on chips look like so turtles all the way down right.
That's literally the exact opposite of my experience with Verilog! The experience of using it feels farther away from the actual hardware than Python does!
No, it's not. It's more like saying a writer should try their hand at different styles and formats of writing. So a novelist who periodically writes non-fiction essays or journalistic content or poetry (whether intended for publication or as a writing experiment), and that practice can in turn inform their writing as a novelist.
Besides, learning programming languages is much easier than learning natural languages. Most experienced programmers today know multiple languages, and most college grads in CS and Software Engineering graduate having used at least 5 or 6 languages if their school is halfway decent.
The language and its concepts is easy as you put it, now taking off the rest of the ecosystem, that is the hard part, which on average based on history, takes around a decade to take off.
If you program just for money, that's one thing. People who only look for money tend to use AI. But if you enjoy programming, that's why you'd pursue exploring languages. And I'd recommend it even to people just looking for jobs as well.
But there are example of "dead" languages whose everyday usage have devolved to only lithurgical functions like: Latin (Catholicism), Pali Cannon (Buddhism) and Sanskrit (Hinduism).
Put Lua in things.
Now, before you anti-Lua types start jumping on that 0-indexed down-vote, let me explain. Its not just about the language, its about the "thing" you're putting it in, its about using the VM effectively, its about planning and using the stack properly. It's about using one language to leverage the strengths - and discover the weaknesses - of other languages, runtimes, ABI's.
Putting Lua in things is easy - and it is hard. It teaches you how to treat other things with respect, and it teaches you how to disregard other things and "Just Use Lua™".
You can learn the language - multiple variants of it. You can learn the 'things' you're putting it in. You can learn how a VM works, what byte code is, how it can be used effectively - and how it can be used counter-productively, too.
It will teach you the limits of a language - Lua, and C - and it will teach you the strengths of both (as well as Rust, and all the other things you can get an FFI scaffolding built around).
You can use Lua to produce your own domain specific language. You can use it to (almost, or indeed completely) replace whatever other language the "thing" is using. You can use it to minimize the impact of a VM and scripting and byte code - you can learn how to maximize these things as well. You can learn byte code and you can leverage that understanding to learn machine language - two ends of a very similar (yet vastly different) spectrum.
You can learn how slow/fast pieces of the system interact productively and effectively - you can learn how things get mightily messed up, also, if you use the wrong lubricant.
In the end, you may not even end up using it for productive things - but you might also do something extremely powerful. (I have.) Either way, by applying this maxim and applying the skills needed to master doing it (or screw it up entirely), you can learn pretty much everything interesting there is to learn about software (besides LLM's) in today's modern stack madness .. and then some.
Seriously, give it a try.
The other one of these developer-tuning activities (there are a few), is to do a Linux From Scratch session at least twice a year. Do it this year, but plan on doing it again next year as well. Do it a few more times this decade.
Iterate. Extend. Ignite.
I put lua in most things I make now. Whether a orchestrator/ansible like replacement or even to support hooks in a mail reader I make or a integration testing style tool - if there is any configuration, I generally pick lua over saddling anyone with another config file that only reflects what the program author could foresee a need for or thought was worth supporting.
Please, if you make tools for others, don’t impose yourself unnecessarily, give people a sandbox- give them lua :)
Another way to do it would be the standard low->high level and then one based on lambda calculus. Von Neumann and Church.
Assembly, C, Java, Clojure.
Assembly, C, C#, F#.
Assembly, Fortran, Rust, Rust, Brainf*ck.
I should have made the list bigger, not smaller. I missed assembly and Erlang.
At the heart of it all it’s just math and logic, and that’s where the different thought studies happen, regardless of which language quirks are navigated.
Both Scheme and CL suffer really badly from a lack of modernization primarily in their stdlibs (or equivalent SRFIs) that tend to make them poor choices for most new professional projects IMO. Great for side projects though. I really wish the next generation of Scheme would take a page from Clojure's book and rework all the SRFIs and bake in immutability as a core requirement.
I made the same problem. I started JavaScript as a C++ programmer. I didn't get prototype based inheritance and kept trying to shovel my C++ style into JS. It took 2-3 years to stop trying to coerce JS into C++ and instead use it as JS.
Now, for many projects I can't stand how I have to write 1000x more code in C++ to insert a debug wrapper around some native API that in JS I can do in a few lines in a few minutes.
Similarly, lisp is great, but it takes a while to really get out to take advantage of its strengths and not treat it like whatever language you're most comfortable in.
When done with the assignment I asked the emacs fan in then lab to look at it and his comment was "you write beautiful C in lisp."
https://www.ee.torontomu.ca/~elf/hack/real-programmers.html
... Besides, the determined Real Programmer can write FORTRAN programs in any language.
---My first revelation with functional programming was when I was dabbling with Clojure based on a Coding Horror question with giving a person being given an algorithim more difficult than Fizzbuzz (which has been memorized)... compute e or pi were the ones that I thought of.
pi / 4 = 1/1 - 1/3 + 1/5 - 1/7 ...
I was trying to figure out if I could use cycle to handle that and then the lightbulb for first class functions hit him.
The solution I came up with was:
(defn pi
([] (float (* 4 (pi 1 0.01 0 true))))
([term tol accum pn]
(let
[t (/ 1 term)
a ((if pn + -) accum t)]
(if (< (* 4 t) tol)
a
(pi (+ term 2) tol a (not pn))
)
)
)
)
The lightbulb was the `((if pn + -) accum t)` bit where `pn` was a boolean that flipped each call... and `(if pn + -)` returned the function.That was May 9th, 2013 (I can find where I was explaining it in a web chat that is still out there).
This happens if the developers don't _want_ to learn. If you're trying out a bunch of languages of your own volition then you won't have this problem.
For prototypes it had Io, it had prolog, erlang for "let it crash" (like elixir).
I feel you're missing the Array languages from your list (APL, J, BQN, ?FORTRAN)
Indeed, new perspectives reveal new ways of thinking.
Instead of hoping you remember to check for NULL, Rust forces you to make a decision on how to handle the possibility of failure. Sure, you can still pick a bad decision (unwrap) but that is at least a specific decision and you can’t forget to do it.
You can take it even further with things like the newtype pattern. If you have lots of numeric IDs (project ID, user ID, etc) you can give them distinct types so you don’t mix them up.
So Rust taught me I can program defensively against my own brain, which sometimes gets tired or lazy or distracted.
It creates an interesting problem, as sometimes you want to know the absence of value. Usually done with the empty "" string.
The other way to handle is that every Object is initiated (pointing) to an starter initial value. (prototype based languages).
Initial values, like null, is an attempt to smuggle the "unset" case into the type without declaring it. And like null, the defensively written logic should prefix every single usage with a check for validity.
Optional is just a natural subset of the generalization that multiple possible states exist; in this case, just the two, and sum types to properly handle bigger sets.
Initial/null never admit more than 2 possible states, and also believe everything is always in one of those 2 states.
Sometimes you do want to represent the concept of “this whole datatype OR nothing”. How do you suggest this is modelled? I think optionality is pretty fundamental. What’s wrong about classical null is that it ascribes optionality to all types regardless of whether they need it.
Being forced to represent optionality with a “special” default is terrible design IMO - see the recent SSH RCE on Musl, largely due to the poor choice of indicating “crypt failed”. Default values can be quite dangerous. Sometimes the domain you are modelling supports such a value, but having “default value” be the default (and not a choice) is a similar evil to default null.
2. Use null coalescing (e.g. `foo?.bar?.baz`)
Languages such as Kotlin and C# support this style o programming.
The thing that I don't like about Optional types (in e.g. Scala or Java) is that they are too verbose. For example, instead of:
return foo?.bar?.baz ?? 0
you have something like: return foo.flatMap(_.bar).flatMap(_.baz).getOrElse(0);
This is even harder to read if you have more complicated logic, requiring implicit lambdas and other machinery.There are caveats, such the function using ? having to return an Option, which can be worked around by using a closure.
I think try blocks will make this slightly easier.
Explicit options nudge you towards a data model that handles the None/null case early, so you don't have to write code that does the coalescing. If you parse-dont-validate early, you more often than not end up with a data model that is simpler and easier to process.
But then, this also depends on the sort of code you write. In the code I write, matching/"if-let"ing on Options to understand their value and then translating that into some enum that encodes the value on the abstraction level of the domain or algorithm has often been really really helpful.
I'm an old developer... I recognise the instinct to resent and avoid new unfamiliar syntax when I've got something that works for me. After some resistance, I give it a go, maybe working on someone else's code that uses it... and then it's suddenly "damn, the lang designers knew what they were doing".
This is one of those. What you might miss is that nested optionality is often the cleanest accurate representation and not error states e.g. a hierarchy of capabilities as a tree of options getting progressively more fine-grained yet you need to work with fine-grained capabilities at the leaves.
Nested optionality appears to be a bad data-structure to avoid when the code it requires is verbose and gruelling. Workarounds to reshape or design for less gruelling code are anti-patterns, adding complexity not to model the problem but to solve a language design weakness, which this solves.
> declared culpability
This is an optional data type, with a different name…
There are even more valid cases where “this integer is not set” and “this integer is set to 0” are different things.
It’s also a type safety concern. Using sentinel values is comparable to the “billion dollar mistake” version of nulls because it creates a special value that implicitly exists and is not part of the variable’s type definition. This means the compiler can’t help you remember to check it. You have to just remember, and forgetting is a bug. With optionals (including opt-in nullability like in Kotlin), on the other hand, the compiler knows when you do and do not have to worry about a value not being present, and can raise a compile time type error when you forget.
That seems so much worse than a `Option<String>`
fwiw john carmack agrees with me
It is a skill issue. But there are only so many skilled programmers. So rust makes sense that even though it made programming a lot boring, you could use droves of unskilled laborers.
There might well not even be one by this standard.
of course you couldn’t help but call rust programmers unskilled couldn’t you. ego.
The ego hit has been mostly a C culture thingie.
You really don't need Rust for that. There are plenty of languages with non-nullable reference-types, and in other languages it's fairly easy to make your own or use a library.
> Rust forces you to make a decision on how to handle the possibility of failure
Let me introduce you to this invention called the ? operator ... also known as the "let somebody else worry about it" operator.
If you mean that errors are not in some global variable that's easy to ignore, then, ok, yes, Rust is not like C in that respect.
Would have also been a much better example for the enum to have the second type being a human with just a single name instead of a cat which is stretching what a person is.
I wholeheartedly agree. I learned this lesson from learning Haskell, before I was even aware of Rust and before Rust had its 1.0 release. One of the reasons I was first excited about Rust in fact was precisely that it had a type system like that of Haskell.
The last time I needed C’s power to fit something into memory and compute limitations, it was code running on a microcontroller and better systems languages like Zig and Rust weren’t a thing yet.
And we already have header-only libraries, which are often close enough.
That is not very accurate. Back then, C was compiled file by file and the outputs from those separate compilations were linked into the final binary (each C file is a translation unit, a term still used today). Header files were reprocessed for each new translation unit that included them, they did not save memory for compilation. They saved some memory on disk storage because shared bits of programs were extracted to a common file, but that savings was also not their purpose.
The purpose was to extract the shared bits and reuse them to prevent errors and reduce duplication. Imagine having to recreate the contents of some or all of stdio.h at the top of all your C files that import it. Now you have hundreds of places to update if that library changed, or hundreds of places to make an error in one of the function specs.
Saving memory was not a benefit, nor an intended benefit of header files. The header still has to be read for each translation unit and processed so there are no memory savings in compilation (back then) from using them. Now, today there can be some with precompiled headers, that come much later than the invention of headers for C, though.
also, the author is a great communicator who genuinely wants to convey his points in clear prose, instead of showing off their intellect (which seems to be the case in some other blogs)
But there are times when I really want to come to Rust from the other direction, and almost think of it as a Haskell alternative that doesn’t bring all the headaches that come from being lazy. The examples the author chose beautifully illustrate some of Rust’s virtues that lend themselves to that perspective.
(But no, we are not going to talk about higher programming. I did say “almost”.)
If you want specific values (integer or otherwise) associated with specific variants (i.e. Human = 1, Cat = 4) you can do that too. Use the #[repr(T)] or impl the From / Into traits yourself.
i.e. type Whatever = ()
That gets you a “tag” that is compile time enforced. You can explicitly match against these aliases and one alias will never == another alias (unless you explicitly write out a from / into impl, and a deref impl to make it happen implicitly).
What you seem to be ultimately asking for is being able to use a tag that is freely floating around to determine the variant of a union. This would need to be explicitly handled inside an unsafe block though.
The “happy path” here explicitly ties them together. Letting them not be handcuffed together is not memory safe, and the compiler has no way to guarantee you’re reading a value correctly.
You can of course bypass all of that inside an unsafe block and pinky promise to the compiler you made no mistakes.
Instead of just getting one type that includes the discriminant and records, Zig and Ada have a tag of some type whose value determines which variant is being used, and then the union (Zig) or record (Ada) is its own type.
I've had it come up when dealing with ASTs, although I don't recall the specifics. This is a problem that's certainly pretty easy to solve with macros. I'm pretty sure last time I ran into this I used `strum` [1] and its `EnumDiscriminants` macro to handle this.
Is there anything you can do with separate types for the tag and the union that std::mem::discriminant and pattern matching can't do?
https://docs.rs/strum/latest/strum/derive.EnumDiscriminants....
The union type always implies a kind, after all, so it's a generalization of the concept.
In Ada, if I use an enum (carrying the example from a link I used in a cousin comment):
type Expr_Kind_Type is (Bin_Op_Plus,
Bin_Op_Minus,
Num);
https://learn.adacore.com/courses/intro-to-ada/chapters/more...Variant records using this enum have to implement the variant for each element of the enum. If I, for some reason, have multiple records using the same discriminant I can know that each one implements it for each value of the discriminant. Does Rust have a way to synchronize your enums similar to that?
enum Parser {
bin_add(...),
bin_sub(...),
...
}
enum Typer {
bin_add(...),
bin_sub(...),
...
}
And I'd have to hope they were both complete and match (at least according to what I know now, that's why I'm asking). If I go back in to the Parser enum and add a `un_minus`, then I have to also go into the Typer enum (apologies for the bad naming, I'm at work and commenting between my way too many meetings) and manually add it. I have no compile time guarantee that both are complete and synchronized. In Ada, I modify the existing enum and both will fail to compile until I add in the variant for `un_minus`. One edit, multiple failures distributed across the code base as a result telling me all the places where I have to add in a new code branch or new variant.EDIT: Run on sentence broken up.
pub trait Payloads {
type Read;
type Write;
}
pub enum FileOp<P: Payloads> {
Read(P::Read),
Write(P::Write),
}
pub struct Inputs;
impl Payloads for Inputs {
type Read = ();
type Write = u8;
}
pub type Request = FileOp<Inputs>;
pub struct Outputs;
impl Payloads for Outputs {
type Read = u8;
type Write = ();
}
pub type Reply = FileOp<Outputs>;
pub fn execute(request: Request) -> Reply {
match request {
Request::Read(()) => Reply::Read(42),
Request::Write(_) => Reply::Write(()),
}
} pub struct Bare;
impl Payloads for Bare {
type Read = ();
type Write = ();
}
pub type BareFileOp = FileOp<Bare>
Correct? So you'd need to create it with BareFileOp::Read(()), but at least it wouldn't carry any overhead of actually containing data.And you don't have to have 1:1 type to enum discriminant:
trait Payloads {
type Read;
type Write;
}
enum FileOp<P: Payloads> {
Flush,
Read(P::Read),
Write { data: P::Write },
}
struct Inputs;
impl Payloads for Inputs {
type Read = ();
type Write = u8;
}
type Request = FileOp<Inputs>;
// enum item construction
Request::Flush;
Request::Read(());
Request::Write { data: 42 };
You can also make an item impossible to exist using the never type: struct WriteOnly;
impl Payloads for WriteOnly {
type Read = !;
type Write = u8;
}I have found it quite useful to be able to generate a dedicated SomeEnumDiscriminants enum for SomeEnum, using the following derive:
https://docs.rs/strum/latest/strum/derive.EnumDiscriminants....
I wrote this on my website in like 2014. I raise it because I think poorly understood problem is usually the reality, and using many languages over time you find ways of describing solutions to those problems and then learning to hone your understanding of the problem.
To me learning new computer languages forces you to think differently about a problem, if you really try to learn the language. I often just write all my side projects in whatever I am trying to learn for as long as it takes from me to learn it. Sometimes I stay with them for very long times if I enjoy them.
That said I look at computer languages like tools. Focus on the right tool for the job. More than once that tool is what others on the project know, because it still solves problems and it’s better that we have a shared toolset to work from.
Guess it also applies to computer programming and hardware description languages.
Spoken languages have the odd idiom that doesn’t translate well, and certain languages express certain ideas more or less succinctly. But it’s not like Chinese’s lack of verb tense makes it harder to think about how things are happening in time. The concept still exists, it’s just expressed differently. But not so differently that I can’t translate decently well on the fly.
Programming languages have much more variability in what kinds of concepts they can easily express. In Smalltalk I’m going to use dynamic dispatch to express damn near everything because it’s just about the only sane way to do domain modeling. In Standard ML I’m going to lean instead on algebraic data types and pattern matching. Those tend to lead to very different ways of breaking down a problem. In Scheme I don’t have either built in but both are easy enough to implement as a library, so I’ll freely choose whichever option seems most appropriate to the problem at hand. In assembly language I’ll avoid any such abstraction altogether because, while I could probably implement versions of both, it would be a total PITA and probably result in incomprehensible code. And translating among any of them is likely to be the work of multiple hours or even days.
But reading this in 2026, when the entirety of the profession has changed (for better or worse) due to LLMs, seems strange.
Like someone writing how some new preservation method for making food last longer revolutionized how they manage their pantry, shortly after the invention of the fridge.
I like the post - it's just sad how many will not have these epifanies anymore, or they wont mean anything, and how many starting today would already just say "how quaint" and move on.
You're just getting older and times change. Your adolescence wasn't any more pure than the kids today.
One could argue you've frittered away your life fiddling bits around instead of becoming a thespian, writer, or musician. Alas, the epiphanies you could have had.
Even in an ideal world, some people would just not want to become artists.
I'm in favor of understanding and knowing how to do all of those things manually - including a healthy does of cursive versus merely digital typing. I'd argue those that can't are incapacitated in a real way.
>One could argue you've frittered away your life fiddling bits around instead of becoming a thespian, writer, or musician. Alas, the epiphanies you could have had.
All those are even more fucked than programmers today.
“A language that doesn't affect the way you think about programming is not worth knowing.” -- Alan J. Perlis.
Yeah. When I tried to code fairly complex thing in Rust, when I was learning it, as I struggled, I simply noticed that I don't have the freedom I'm used to when it comes to at which level of my architecture I create objects or perform writes or who can pass a reference to whom. Once I learned I can't do it, and not that I just don't know how, suddenly things had their proper place. Yes, I don't have full freedom how I architect things here, but until the end of the project, I haven't found a single case where this restriction would lead to worse not better architecture. Once I leaned into it, as the author says, later I caught myself many times thinking, that C++ would happily let me do it which would result in a bug that would be super hard to find, that would probably burry the whole project.
It's just another tool not a philosophy of life or benchmark of intellect.
This is happening in Python land as well. now that type annotations are a normal everyday thing we are doing a lot of code generation. Of course I'm sure how Rust does it is much better integrated and "correct" and all that but the Python thing of generating functions/classes/etc. on the fly is not quite the sole choice anymore for various reasons, and pep484 and related are making static code generation more feasible.
Not really. There are two different macro systems, one of them has you writing cryptic functional style pattern matching on expressions and the other just gives you a stream of tokens to parse yourself.
Neither of them lets you create types, functions or traits as objects like in Zig. You need to build Rust expressions that produce them and keep track of all the names while doing it. If the macro fails to put out valid code the user gets a confusing error, unless you make them use the nightly compiler. It's one level above C style string manipulation.
And did I mention that if your macro extends Rust syntax you need to pull in a third party Rust parser, because the one in the compiler is strictly internal?
I do want to push back on one point though:
> It also stops you from mutating data you're also reading elsewhere
No it doesn't. It stops you from mutating data _that it controls_ which you're also reading elsewhere _also within your Rust code_. But Rust presumes to be a systems programming language. In a system, there are other processes, other devices, interaction across the network, etc. Any of those breaks the integrity of Rust control, and forces you to contend with other independent entities which allocate or deallocate resources, write when they feel like it etc.
> making it impossible to stumble into whole classes of weird undefined behavior, especially with code that runs in parallel.
... and this is the half-false promise that belongs in marketing talk. Although if OP means "undefined behavior in the sense of "undefined by the language spec", that's a whole other kettle of fish.