I'm a little worried on behalf of the "Python Language Tooling Team" at Meta, because uv has been so popular, and I wouldn't be surprised if ty wins out in this space.
So watch out, or this will become like Atom or Flow, an internal competitor of a technology that is surpassed by the more popular external open source version, leaving the directors/vps muttering to themselves "It's too bad that this team exists at all. Could we get rid of them and just switch to the open source stuff?"
Perhaps just something for the manager (Aaron Pollack?) to keep an eye on....
Hey Kevin, we overlapped for a bit during your time at FB when I was working on Flow. Nice to hear from you!
I’m working on Pyrefly now, but worked on Flow for many years before. For what it’s worth, we are taking a different approach compared to Flow and have explicitly prioritized open source and community building, something I know we both care a lot about.
Of course, nothing is guaranteed and we’ve seen plenty of volatility in bigco investments to infra lately, but I do believe we’re starting this journey on the right foot.
Meta seems to place a pretty high premium on controlling its open source projects, especially dev tooling. I guess dating back at least to the git maintainers telling them they were doing things wrong with their monorepo and refusing to upstream scale fixes, which precipitated their migration to mercurial (who were more than happy to take the contributions).
Given the change velocity of internal tooling you can understand why owning your own project makes sense here.
Hi folks, I work on the Pyrefly team at Meta. Our FAQ covers a good number of the questions raised here: https://pyrefly.org/en/docs/pyrefly-faq/. I can also try to answer some of your questions. Thanks for taking a look!
Yes. If you want runtime validation of data you’re taking in people recommended pydantic. If you’re looking for runtime validation within your own code I’ve seen people use beartype, though to be honest I don’t personally understand the value added from it
"Today we’re releasing Pyrefly as an alpha. At the same time, we’re busy burning down the long-tail of bugs and features aiming to remove the alpha label this Summer. "
Why is "written in Rust" a feature to be mentioned? Who cares? So my type checker has memory protection and is compiled. I'm not running my type checker on an embedded system or in a mission critical service. It seems kind of like "written in Erlang". I'd prefer to have non-performance critical code for Python written in Python. That way the broader community can support and extend it.
Have you used Rust before? As a user, the speed and safety is nice. But as a developer, Rust projects are easier to hack on and contribute to.
That's kind of the whole appeal of Astral. I know Python better than Rust, but it's a lot easier for me to hack on Rust projects. The whole appeal of Astral is that they want to bring Rust-quality tooling to Python.
> Rust projects are easier to hack on and contribute to.
This was actually the subject of a study at the University of Waterloo:
> We find that despite concerns about ease of use, first-time contributors to Rust projects are about 70 times less likely to introduce vulnerabilities than first-time contributors to C++ projects.
> Rust projects are easier to hack on and contribute to.
You can say that about any languages that you yourself know, or other people know. There are beautiful codebases in many other languages, and awful ones in the same languages.
If your Rust codebase has a lot of unwraps and lifetime annotations (among other things), I will probably not find it a joy to contribute to it, at all.
Rust has very arcane syntax and a lot of rules that developers coming from interpreted / garbage-collected languages (like the ones using these tools) would have a hard time grasping. It’s easy for people who are already familiar with it, but isn’t that always the case?
An LSP is performance-critical code. It directly affects responsiveness of your IDE, or even the viable scope of a project that the LSP can handle.
Rust is both CPU- and memory-efficient, quite unlike Python. (It could have been OCaml / Reason, or Haskell, they are both reasonably fast and efficient, and very convenient to write stuff like typecheckers in. But the circle of possible contributors would be much narrower.)
The circle of possible contributors doesn't really matter. It's a Meta project, they have others written in OCaml and to this day they manage to have contributors eg https://github.com/facebook/flow because they hire and pay them.
To be fair, pyright's performance as an LSP is tolerable. The main issue to me is running the type checker as a pre-commit hook or as part of CI. Mypy is awful, though.
> I'd prefer to have non-performance critical code for Python written in Python
A type checker is performance critical code. You can watch how Pylint, just a linter, written in Python, lints your source code line by line. It's so slow it can take 30 seconds to update the linting after you change some lines.
It makes it easy to find performant and quality software by searching for "[insert tool description] rust", I personally don't mind! Seeing how 95% of the tools I use on the daily are written in Rust, I love finding new ones and am rarely disappointed.
>Why is "written in Rust" a feature to be mentioned? Who cares?
A lot of people. Correct or not, I think "written in rust" have become synonymous with "very much faster than the alrernatives" in the pyrhon community.
I feel like the likelihood that a project will say what language it is written in is much higher if that language is Rust. I like Rust but I do find this trend a little annoying. (Even though I acknowledge that "written in Rust" probably means the tool is relatively new, not buggy, and easy to use.)
The Rust code written here is so easy to follow but all these new Python tooling being written in Rust worries me, it adds yes another vector to the N-language problem.
For the Python ecosystem, it's natural to use Python where Python can cope, and a high-performance language where it cannot. There are two such languages in wide use around Python: Rust, and, inevitably, C. So N = 3.
(C, to my mind, should be eventually phased out from application programming altogether, so N would be 2, but it's a loooooong process; Python may become a legacy language before it converges.)
Yes but the idea is that by slightly upgrading python code to mojo (which is a controlled superset of python), you get complied very high performance code. So for example if it were possible to convert mypy to mojo it could be as fast as rust but pythonic.
Sadly I think the ship has sailed and Rust has hit critical mass now. Personally I find it aesthetically awkward, but for Python integration and tooling it seems like Rust has become the default C replacement. You would think Python devs might have preferred something more superficially Pythonic like Nim or perhaps something more C-ish like Zig, but those projects don't have the same buzz so here we are. There's probably more young devs who are into Rust than C nowadays.
I am not holding out much hope for Mojo because it feels deeply embedded in the AI/LLM hype space instead of being presented to Python devs outside of that niche as a useful language extension in its own right.
I don't think it really matters whether Rust has hit critical mass or not tbh, just the fact that it is entirely a new language to learn with very different semantics compared to Python is a blocker for many people.
Mojo right now is not much better, but I've seen Python compatibility factor into the language design and semantics again and again. It is not enough to be a language that looks like Python, like Nim, things also have to behave the same when the semantics of static typing allows.
Mojo is not deeply embedded in the AI/LLM hype, there is nothing in the language that is targeted specifically for AI. The standard Library has a GPU package for general-purpose gpu programming, but that isn't AI specific.
I just tried pyrefly on a project that really needed it. It complained about an assignment of a new value to a global int variable within a function, even though the function contained the 'global' statement that should have made that OK, I think. I know that globals and assigning to them here and there are problematic for real good software, but I am surprised that Pyrefly is stricter than python on something that I don't see as a type- checking issue. But it did find a decent list of other problems that I haven't finished working my way through.
I had gotten so messed up trying to put together a quicky hobby-type program to create a data structure of perhaps a hundred data items in various overlapping and inter-related hierarchies, tuples, dicts, and lists akimbo, that I gave up on it about 10 days ago. I hypothecated that bondage and discipline might be the way to control the confusion, so I'm rewriting, using SQLite for the dataflow from function to function, lots of little tables and no optional fields. Can anyone opine on whether that is a sensible option?
Thanks for trying it out! If you run into any blockers, please let us know by filing a GitHub issue or sending us a message on Discord. Pyrefly is still alpha software, so bugs are expected, but your feedback is extremely valuable as we work to squash them.
Readit News
So watch out, or this will become like Atom or Flow, an internal competitor of a technology that is surpassed by the more popular external open source version, leaving the directors/vps muttering to themselves "It's too bad that this team exists at all. Could we get rid of them and just switch to the open source stuff?"
Perhaps just something for the manager (Aaron Pollack?) to keep an eye on....
I’m working on Pyrefly now, but worked on Flow for many years before. For what it’s worth, we are taking a different approach compared to Flow and have explicitly prioritized open source and community building, something I know we both care a lot about.
Of course, nothing is guaranteed and we’ve seen plenty of volatility in bigco investments to infra lately, but I do believe we’re starting this journey on the right foot.
Cheers, Sam
Given the change velocity of internal tooling you can understand why owning your own project makes sense here.
Deleted Comment
Did you see better speedups over Pyre? Aka did I leave stuff on the table…
I didn’t know about the Rust-based Hack checker— that’s really cool!
Deleted Comment
Deleted Comment
"Today we’re releasing Pyrefly as an alpha. At the same time, we’re busy burning down the long-tail of bugs and features aiming to remove the alpha label this Summer. "
That's kind of the whole appeal of Astral. I know Python better than Rust, but it's a lot easier for me to hack on Rust projects. The whole appeal of Astral is that they want to bring Rust-quality tooling to Python.
That static typing is nice, I wonder if it's going to catch on one day.
The amount of energy spent trying to bend dynamically types languages into being real ones is just comical. Even the standard library is barely typed, so they give no fucks https://github.com/python/cpython/blob/v3.13.3/Lib/re/__init...
What does it accept? Who knows. What does it return? Don't worry about it
This was actually the subject of a study at the University of Waterloo:
> We find that despite concerns about ease of use, first-time contributors to Rust projects are about 70 times less likely to introduce vulnerabilities than first-time contributors to C++ projects.
https://cypherpunks.ca/~iang/pubs/gradingcurve-secdev23.pdf
You can say that about any languages that you yourself know, or other people know. There are beautiful codebases in many other languages, and awful ones in the same languages.
If your Rust codebase has a lot of unwraps and lifetime annotations (among other things), I will probably not find it a joy to contribute to it, at all.
Rust is both CPU- and memory-efficient, quite unlike Python. (It could have been OCaml / Reason, or Haskell, they are both reasonably fast and efficient, and very convenient to write stuff like typecheckers in. But the circle of possible contributors would be much narrower.)
A type checker is performance critical code. You can watch how Pylint, just a linter, written in Python, lints your source code line by line. It's so slow it can take 30 seconds to update the linting after you change some lines.
Open source Rust is still review able.
A lot of people. Correct or not, I think "written in rust" have become synonymous with "very much faster than the alrernatives" in the pyrhon community.
Deleted Comment
If one is a "purist", the idea of non-python tool involvement may dissatisfy.
Scare-quoting "purist", given the general lack thereof anywhere in Open Source, python itself being a case in point.
I hope Mojo can offer something here
(C, to my mind, should be eventually phased out from application programming altogether, so N would be 2, but it's a loooooong process; Python may become a legacy language before it converges.)
I am not holding out much hope for Mojo because it feels deeply embedded in the AI/LLM hype space instead of being presented to Python devs outside of that niche as a useful language extension in its own right.
Mojo right now is not much better, but I've seen Python compatibility factor into the language design and semantics again and again. It is not enough to be a language that looks like Python, like Nim, things also have to behave the same when the semantics of static typing allows.
Mojo is not deeply embedded in the AI/LLM hype, there is nothing in the language that is targeted specifically for AI. The standard Library has a GPU package for general-purpose gpu programming, but that isn't AI specific.
I had gotten so messed up trying to put together a quicky hobby-type program to create a data structure of perhaps a hundred data items in various overlapping and inter-related hierarchies, tuples, dicts, and lists akimbo, that I gave up on it about 10 days ago. I hypothecated that bondage and discipline might be the way to control the confusion, so I'm rewriting, using SQLite for the dataflow from function to function, lots of little tables and no optional fields. Can anyone opine on whether that is a sensible option?