Rewriting in Rust

This feels like a really good blog post to write - something I'd really want to read - but the generated AI prose distracts, and makes me lose trust and not want to read this. I'm very AI-pilled, so this is not a criticism of AI as a whole, just this piece.

Note, I am a co-maintainer of GNU coreutils. Whether that makes my opinion relevant, biased, or both, you can decide. :)

I really wished the documented their benchmarking methodology here, or at least cautioned the reader not to jump to conclusions based on the benchmarks shown.

GNU 'sort' performance can drastically be altered by the locale in use, the input, and the arguments given to the --buffer-size and --parallel options. GNU 'sort' is fairly conservative in how many threads it will use by default, and in my experience, much more so than uutils. This is because throwing more threads at 'sort' may make it faster (or may not), but also risks running out of memory. This is an issue with uutils, which is poor at deciding when to use external sorting:

  $ export LC_ALL=C
  $ for i in {a..z}; do yes $i | head -n $(numfmt --from=iec 512M) | tr -d '\n' >> input; done
  $ time sort input > /dev/null

  real 0m24.245s
  user 0m0.896s
  sys 0m19.161s

Here is the same command using the latest uutils commit compiled with 'make PROFILE=release':

  $ time uu-sort input > /dev/null
  Killed                     uu-sort input > /dev/null

  real 2m53.560s
  user 1m40.634s
  sys 0m59.847s

The process gets killed by the OOM killer. This is likely because uutils 'sort' decides to use 18 threads, instead of the 1 used by GNU 'sort'. I find it a bit frustrating that benchmarks are thrown out without any methodology or citations, because they are often trusted without question. These could be benchmarks from before uutils had localization, which was the case before 2025, and treated LC_ALL=en_US.UTF-8 as LC_ALL=C. In that case, of course it would be much faster than GNU coreutils, but it also means uutils would give you the wrong results for non-ASCII characters. There is, as shown above, much more considerations beyond speed that seemingly never get the time of day next to flashy benchmarks...

For a brief second I thought JetBrains were rewriting their tools in Rust and we were going to get some performance improvements.

This post advocates rewriting incrementally instead of all at once. Just like Joel said back in 2000, and like everyone continues to always say to this day. Yet in practice, people don't actually do this; complete rewrites in Rust remain far more common than incremental ones, particularly when rewriting from a language other than C. The high-profile exceptions, like Linux, Windows, and Firefox, are codebases so huge and ancient that they obviously cannot be rewritten from scratch. When rewriting from scratch is an option, it tends to be taken.

The reason for this is pretty straightforward: Incrementally porting a codebase from another language to Rust (especially if it's not C) is a deeply unpleasant experience, because the interop tooling isn't good enough and you spend most of your time fighting it. Consider the case of rewriting from C++; in the simplest case, you use bindgen and cbindgen, which only work with extern "C" functions in both languages. So you effectively have to rewrite each API first from idiomatic C++ to C-in-C++, then translate to C-in-Rust, then rewrite again in idiomatic Rust. And then repeat for the next API. And so on. It's not going to take long for most programmers to go "screw it, I don't care what Joel said, at least when I rewrite all my work I'll be doing it in one language where anything can call anything else". cxx and autocxx modestly improve things, but still leave you with an impoverished API vocabulary and similar problems, and you still have to do the three-step rewrite for each API, one at a time.

This is also why TypeScript, Kotlin, and Swift worked so hard to have seamless two-way interop with JavaScript, Java, and Objective-C. Without that, they couldn't have credibly promised to replace the earlier languages (because large existing codebases where a rewrite wasn't economical would still be stuck with them), and so couldn't have gotten off the ground.

Crubit is supposed to fix this for C++-to-Rust, and I'm rooting really hard for it, but it's not there yet. For most other languages, a Crubit-like thing probably isn't even possible in principle, because the differences are too great.

(I'm not talking here about the use case where you started with a garbage-collected language but have hit a performance ceiling with it, so you rewrite just the most performance-sensitive parts in Rust, while continuing to develop the rest of the codebase in the original language. This is often a great way to use Rust, but it's solving an easier problem and so poses fewer difficult tradeoffs.)

Small related anecdote:

Was just testing latest gen LLM capabilities, and decided to give it goal of rewriting a small opensource project in Rust. (Should be noted: was not some tiny library, but an actually useful network service).

It completed the entire rewrite from typescript to rust in about 2 hours, ~600k tokens used. Worked perfectly on first try with no follow up changes required. Memory and CPU usage now a tiny fraction of TS version (obviously). Rust code was simple, easy to read, accurate test suite, etc.

I was pleasantly surprised.

Obviously bigger code bases with more complex business logic will likely struggle here, but there are some advantages to "RIIR" when performance matters, even security benefits aside. Rust can help squeeze more juice out of old hardware; reduced memory footprint especially helpful with current RAM prices.

For small services where operational cost matters, having LLMs "rewrite it in rust" might be worth the spend.

This topic has somewhat recently acquired a connotation where it yields a polarized response. People seem "tired" of the language and the episodes like Bun rewrite don't help, especially in current LLM-obsessed era.

Amidst this, just let me try to give my own experience with language while working in a field where it doesn't have much traction (scientific/numerical programming). A couple of years ago, after being tired trying to make Python faster, I was looking for a language to write simulation code in. Fortran (and C/C++) has been the go-to choice in my field for decades but I wanted to try a modern language. I tried Julia but the workflow didn't come naturally to me. Also by default, it also wasn't ahead-of-time compilable. Rust was my second choice but it's type system/design, expressiveness and tooling (rust-analyzer) won me over. I have written so much of it and I really enjoy writing it. Borrow checker isn't really an issue 99% of time when writing scientific code. And when it is (self-referential data structures), you can just use an arena (or something equivalent). The C/Fortran inter-op is great so it's still so easy for me to build my own abstractions on top for the more general stuff like solving ODEs and sparse matrices.

And of course, there is speed, memory safety and ability to parallelize things so effortlessly (rayon is magical), but I feel those were never the things that actually won me over. The language was just a joy to write.

I would love it if jetbrains were to speed up IntelliJ. I started a new job writing Java professionally last year and, up that point, I had no idea that IDE's could actually be that slow.

My company gave me a brand new, beefy Macbook pro and that thing can barely handle IntelliJ sometimes...

That comparison figure of LaTeX vs Typst is ridiculous, choosing math-dense example for LaTeX vs classical markdown text for Typst. Yes, Typst syntax is much simpler https://typst.app/docs/reference/math/ but one doesn't need to make unfair comparisons to make Rust look better, Rust has a lot of good things going for it anyway, for example cargo.

I think he forgot to praise cargo.

(Imho, having cargo available is often already worth a RIIR.)

I think that a line-for-line rewrite in Rust like Bun did isn’t really getting all of the benefits of using Rust. A key benefit is utilizing the type system to make the borrow checker work for you by making certain checks happen at compile-time instead of runtime. Making it impossible for certain mistakes to occur is a massive benefit but you need to restructure everything to do so.

To me Rust is great where you don’t have to touch unsafe part. All the hard works are done by those skillful maintainers of the language and some core libraries such as tokio. I don’t see much value in it if either I have to work on the low-level stuff quickly (like implementing linked lists myself) or a complete high-level of i/o bound work like writing web servers.

I understand that's a guest post in Jetbrains blog, and these guests could very well be real people, project maintainers, conference speakers and whatnot, but I feel they used LLM so heavily that it reads like a slop. Pease do better next time.

I think every project should seriously ask itself if it actually needs manual memory management. My hunch is that most people who think they need it, do not. If you really truly do need it, then yes Rust is a great way to get most of the benefits of garbage collection while still having manual control over memory.

> ... Part of this comes from auto-vectorization: the Rust compiler generates SIMD instructions automatically from a regular for loop, while most C implementations require hand-written SIMD...

I understand that their compiler in Rust is better. And any other language could compile it using the same SIMD instructions.

Well they should rewrite their shitty sluggish new JS UI in Rust then.

> our attempt to give it an honest look.

Oh no, what are we hiding?

> introduces bugs you already fixed.

Isn't this why every bug fix gets a unit test?

I appreciate rewrite isn't always the answer but it's a strange post when the authors should be giving constructive ways forward so we move to Rust and then use their 1k star Django clone they link to.

Instead anytime a question is posed it falls back to "sometimes" and little detail after that.

Citing a nine year old paper instead of something newer or doing their own benchmarking wasn't ideal either.

> And then there’s the Stack Overflow Developer Survey.

Honest question, at what point is the Stack Overflow Developer Survey not representative of the average software engineer, many (most?) of who no longer use Stack Overflow?

> Typst is a good example of the readability argument:

I'm afraid the author completely missed the point here. Typst is not more readable than (La)TeX because it's written in Rust, but because its DSL was so designed. It could have been implemented in PHP that the result would be exactly the same.

Same thing with the auto-unroll/SIMD arguments. AFAIK, it's LLVM that's doing the job, there is nothing theoretically preventing a {language} compiler to obtain the same results.

Where Rust shines w.r.t. other similar languages though is that its strict memory model lets developers push further memory/concurrency optimization without sacrificing safety – which she highlights in the GNU/coreutils sort comparison.

And here I thought they’re finally rewriting their slow, bloated IDEs. Shame.

"Typst instead of LaTeX"

Is this really an alternative? I would never replace battle proved TeX with consistent syntax, build for processing text, with great fonts, thousands of plugins for some Markdown mess, which is, in addition, paid.

Only because it is written in Rust, not in C.