Hi Hackernews, I’m on the team working on Pyrefly at Meta. We address a lot of the comments in our FAQ: https://pyrefly.org/en/docs/pyrefly-faq/. I will try to address some questions directly too!
Pyrefly is a work in progress, but you can install it from pypi and/or try our alpha VSCode extension if you are interested. You may still get false positive errors on your project as we are still burning down bugs and some features. We would love comments, bugs, or requests on our GitHub as you try it out.
Lastly, one of our team members is giving a talk at PyCon about the design if folks are interested: https://us.pycon.org/2025/schedule/presentation/118/. We didn’t expect much attention before then, so thanks for taking a look!
On one hand, "launching Spring 2025". On the other hand, 47% complete, with a ton of basic stuff not yet ready. I wonder how fast they plan to move, given that only 32 days remain of the spring.
The astral team (behind uv and ruff) is already working on a type checker in Rust, and given the quality of their existing tools, I'm inclined to wait and see what they release. Pyrefly looks interesting, but from the repo it seems pretty early-stage and not intended for external use yet.
Also Facebook has a history of releasing the source code for something. Making a huge splash and then essentially doing nothing after 3 months (aka after someone gets their review).
They use the code internally but fail at making sure it has use externally. This is doubly the case for anything infrastructure
Buck2: Was released. Never could be built correctly without the latest nightly of Rust and even then it was fragile outside of Meta's build architecture
Sapling: Has a whole bunch of excitement and backing when it was announced. Has been essentially dead 3 mos after release.
I used to work for Meta infra. I know the MO. Have a hard time trusting.
Astral use-case is external and has a better chance of actually being supported.
We know we can't just ask for trust upfront. Instead, we want to earn it by showing up consistently and following through on our commitments. So, take us for a spin and see how we do over time. We're excited to prove ourselves!
Sorry I didn't mean it's not dead but it really hasn't got as much feature support. Things like LFS support got deproritized just because the internal team asking for it got a different feature.
Both are EXTREMELY active but only for the needs of Meta and not for the community.
Adoption outside of Meta is nearly non-existent because of this.
Look at something like Jujitsu. instead of Sapling and you can see a lot more community support, a lot more focus on features that are needed for everyone (still no LFS support, but it wasn't because Google didn't need it).
I guess I don't consider a larger number of commits as actively supporting the community. The community use is second place and the open source is just a one time boost to recruiting PR.
When I was there (which was a while ago) almost every decision was based around PSC (Performance Summary Cycle) and it's easy to justify a good rating for a large project being open sourced. Less so to make sure it's well supported for the use cases of the community.
It has warnings that steer you away from idiomatic Python towards less idiomatic Python.
Idiomatic Python favours EAFP (Easier to Ask Forgiveness than Permission) over LBYL (Look Before You Leap), so you might write something like this when using TypedDicts:
Pyright warns you about this and pushes you to write this:
if "baz" in bar and "qux" in bar["baz"]:
foo = bar["baz"]["qux"]
...
else:
...
They identified this as a bug and fixed it, then changed course and decided to reinstate the behaviour. So if you want to write idiomatic Python, you need to disable typechecking TypedDicts.
Now if this were a linter then I would understand. But a type checker should not be opinionated to the point of pushing people not to use idiomatic Python.
Nice, I didn't know about that lib! Usually if I have very nested content I recursively parse it and I'll have some business logic in the function to handle cases here and there. Or I'll have a validation library like Pydantic.
When I last compared it to mypy a few months ago adding typechecking to an old project that had types but I had for some reason never actually run a typechecker on it:
* Was overwhelmingly slower than mypy
* Had a few hundred more false positives. I gather from reading their philosophy afterward that this was on purpose. Rigid dogma > doing the right thing in the circumstance in their opinion.
* Did not find any actual bugs, whereas mypy identified 3 errors that lead to fixing real issues AND had fewer false positives, due to its better understanding of python code.
* Comically overweight with its typescript dependency.
My first impression of it was of a very low quality, over engineered project prioritizing noise over signal. Looking forward to trying out the astral typechecker as well.
It's slow as hell. And every version bump comes with a bunch of new warnings/errors by default that is quite disruptive if you want to keep it quiet. But mostly that it is annoyingly slow.
I guess it’s all relative. One of my code bases still has pylint running with only a couple custom lint rules, that one is slow as hell.
As for version bumping, maybe it’s just a me thing, but I hard fix a version and only update occasionally. Sure each update brings new warnings, but most of them are valid and if you only do it a couple times a year… not that big a deal.
Try ruff instead of pylint. It will open your eyes to how performant software can actually be. All of us that have been working in python and JavaScript for too long have really forgotten.
We use ruff for everything that isn’t a custom rule.
The only thing I found it ‘missing’ was an indentation check (it’s in preview, I don’t turn on preview rules), but I realized it doesn’t matter because we also have a formatter running on everything.
pyright doesn't do well with inspecting inherited types, at least in our experience. Falls far behind JetBrain's type checker in Pycharm. (This is one reason why pycharm >> zed.)
Are you kidding? Pycharm's pychecker is very incomplete. My whole team moved away from pychram due to the shitty typehint support which barely moved forward. (5+ year old bug, inconsistent behavior across the different features in pycharm indicating to me a fragile mess of heuristics behind the curtains)
I agree that most of the functionality on top of the type checker/engine is better in pycharm though. But a correct implementation is more important
At least the alpha milestone is May 12, but I don't think they've publically committed to anything.
In particular for me it would be great to have a better typechecker than Pyright available for Zed (basically why the editor is a nonstarter right now).
An editor is a nonstarter because it doesn't doesn't integrate the right implementation of some particular optional feature for some particular language?
I have some sympathies with the parent post as Zed is my daily driver and I mostly write Python. Pyright is horrible to work with, tons of false positives. Parent poster - you might find this useful: https://github.com/zed-industries/zed/discussions/24801
Well, I mean why use a suboptimal tool? I don’t have any ideology to follow, my codebase is just mostly Python and heavily typed. Pyright doesn’t support all the features so I stick with VSCode/Pylance.
So what editor is a starter then? As far as I know pyright is the de facto checker for python in most editors. Not sure why it would be worse in zed, seems to work fine for me. Or as well as pyright can work.
I'm wondering though, why do existing tools like Mypy and Pyright not scale for Meta? Is Pyre(fly) being used extensively at Meta? In general what sort of issues did Meta run into?
We go into it a bit more here: https://pyrefly.org/en/docs/pyrefly-faq/ The short answer is that the scale of the Python projects we work on is huge. Pyre is the only checker that can scale for our needs today. To bring modern features like type inference and better IDE integration, a full rewrite needed to happen. We are not done just yet, but have been using our massive codebases to test the performance and flexibility of the checker. Thanks for the question!
It's hard to tell. Other similar options in other commands can mean things like "exit with status 0 even if errors are detected". Without installing pyrefly, I'm not sure what that flag's supposed to do. It's not documented on their site.
It adds a comment to suppress the error above each line in your codebase that causes an error, but it reports the errors before doing that, so you can only see the result after you re-run `pyrefly check`.
Oh! That's nifty. It leaves you with a baseline so you can stop making new mistakes, and gives you something to grep for when you want to fix existing ones.
Thanks for giving this a try! We go over how to use the upgrade tooling (`--suppress-errors`) in the guide here: https://pyrefly.org/en/docs/installation/. This is intended to make upgrading from different versions of type checkers easier. It’s something we use internally and wanted to share with the community along with the checker itself.
We also allow you the ability to suppress whole classes of errors if you want to ignore specific error types and avoid inline ignores: https://pyrefly.org/en/docs/configuration/ A word of caution: this will ignore future errors of this type as well.
First, make sure your project is properly configured and then follow the instructions. `--suppress-errors` will add ignores inline allowing the project to check cleanly. We have thought about a feature that allows you to keep suppressions in a separate file, but we have not put it on the roadmap yet. If this is something important for your workflow, we would like to learn more - please add a feature request on GitHub.
I expected that I would be able to run the check command and it would just work. Upon reading the docs, this tool recommends incremental adoption, and after using `--suppress-errors`, `# pyrefly: ignore` is all over my codebase.
I know I shouldn't get into this thread, but I'm extremely curious: what did you expect should have happened? I mean, literally, what do you mean by "just work", what work did you expect a type checker to perform if not to show you errors?
I think the point here is that for something like Python the default behaviour should be an assumed return type of `Any` rather than throwing an error. Maybe that is the case and the GP had configured it otherwise.
No it shouldn't because then you can easily miss places where you should have added a type annotation, and also your lazy colleagues won't bother adding them at all.
The worst of both worlds is colleagues who insert type annotations but they're wrong. And there's no CI pipe to verify the types before check-in.
Working with these things makes me often think that people who want to write high quality software should just use a better language. But, well, real world is real.
Yeah unfortunately Python is crazy popular so a lot of the time you don't get a choice.
If fairness if you set up uv and Ruff and Pyright it's kind of ok. Still have to deal with the general noobness of the ecosystem and the horrific performance, but I've seen worse.
Hi Hackernews, I’m on the team working on Pyrefly at Meta. We address a lot of the comments in our FAQ: https://pyrefly.org/en/docs/pyrefly-faq/. I will try to address some questions directly too!
Pyrefly is a work in progress, but you can install it from pypi and/or try our alpha VSCode extension if you are interested. You may still get false positive errors on your project as we are still burning down bugs and some features. We would love comments, bugs, or requests on our GitHub as you try it out.
Lastly, one of our team members is giving a talk at PyCon about the design if folks are interested: https://us.pycon.org/2025/schedule/presentation/118/. We didn’t expect much attention before then, so thanks for taking a look!
A rewrite of https://github.com/facebook/pyre-check in rust?
Yes, this is a ground up rewrite: https://pyrefly.org/en/docs/pyrefly-faq/#what-is-the-relatio...
My thought exactly! Maybe for performance reasons? (Though I wouldn't expect OCaml to perform that badly, either…)
On one hand, "launching Spring 2025". On the other hand, 47% complete, with a ton of basic stuff not yet ready. I wonder how fast they plan to move, given that only 32 days remain of the spring.
Not in southern hemisphere.
It's not "47% complete", it's just that (now) 52% of the issues in a milestone are complete.
We’re 53 days away from summer though.
The astral team (behind uv and ruff) is already working on a type checker in Rust, and given the quality of their existing tools, I'm inclined to wait and see what they release. Pyrefly looks interesting, but from the repo it seems pretty early-stage and not intended for external use yet.
Also Facebook has a history of releasing the source code for something. Making a huge splash and then essentially doing nothing after 3 months (aka after someone gets their review).
They use the code internally but fail at making sure it has use externally. This is doubly the case for anything infrastructure
Buck2: Was released. Never could be built correctly without the latest nightly of Rust and even then it was fragile outside of Meta's build architecture
Sapling: Has a whole bunch of excitement and backing when it was announced. Has been essentially dead 3 mos after release.
I used to work for Meta infra. I know the MO. Have a hard time trusting.
Astral use-case is external and has a better chance of actually being supported.
We totally get why you might be skeptical and even address this in our FAQ: https://pyrefly.org/en/docs/pyrefly-faq/#how-do-i-know-this-...
We know we can't just ask for trust upfront. Instead, we want to earn it by showing up consistently and following through on our commitments. So, take us for a spin and see how we do over time. We're excited to prove ourselves!
Sapling is actively developed, not "dead after 3 months": https://github.com/facebook/sapling/commits/main/
Have not tried building Buck2 (no personal use for it), but its also actively developed: https://github.com/facebook/buck2/commits/main/
Sorry I didn't mean it's not dead but it really hasn't got as much feature support. Things like LFS support got deproritized just because the internal team asking for it got a different feature.
Both are EXTREMELY active but only for the needs of Meta and not for the community.
Adoption outside of Meta is nearly non-existent because of this.
Look at something like Jujitsu. instead of Sapling and you can see a lot more community support, a lot more focus on features that are needed for everyone (still no LFS support, but it wasn't because Google didn't need it).
I guess I don't consider a larger number of commits as actively supporting the community. The community use is second place and the open source is just a one time boost to recruiting PR.
Is this standard promotion driven development? Or do the people who are trying to open source these products end up being blocked?
When I was there (which was a while ago) almost every decision was based around PSC (Performance Summary Cycle) and it's easy to justify a good rating for a large project being open sourced. Less so to make sure it's well supported for the use cases of the community.
It is codenamed "Red Knot".
Refer to the crates starting with red_knot in this directory to follow its development: https://github.com/astral-sh/ruff/tree/main/crates
The latest commit was only an hour ago.
astral team has live web demo they use internally of redknot on https://playknot.ruff.rs
At first I thought this was the release from Astral for their type checker. Same boat, sitting tight, holding champagne ready for that day too.
Cannot wait to get rid of pyright, but I’ll be holding out for Astrals type checker.
I've loved pyright so far, what do you dislike about it?
It has warnings that steer you away from idiomatic Python towards less idiomatic Python.
Idiomatic Python favours EAFP (Easier to Ask Forgiveness than Permission) over LBYL (Look Before You Leap), so you might write something like this when using TypedDicts:
Pyright warns you about this and pushes you to write this: They identified this as a bug and fixed it, then changed course and decided to reinstate the behaviour. So if you want to write idiomatic Python, you need to disable typechecking TypedDicts.— https://github.com/microsoft/pyright/issues/1739
Now if this were a linter then I would understand. But a type checker should not be opinionated to the point of pushing people not to use idiomatic Python.
For maximum safety, I would write:
any more than 2 levels deep of this pattern and I reach for python-benedict
Nice, I didn't know about that lib! Usually if I have very nested content I recursively parse it and I'll have some business logic in the function to handle cases here and there. Or I'll have a validation library like Pydantic.
When I last compared it to mypy a few months ago adding typechecking to an old project that had types but I had for some reason never actually run a typechecker on it:
* Was overwhelmingly slower than mypy
* Had a few hundred more false positives. I gather from reading their philosophy afterward that this was on purpose. Rigid dogma > doing the right thing in the circumstance in their opinion.
* Did not find any actual bugs, whereas mypy identified 3 errors that lead to fixing real issues AND had fewer false positives, due to its better understanding of python code.
* Comically overweight with its typescript dependency.
My first impression of it was of a very low quality, over engineered project prioritizing noise over signal. Looking forward to trying out the astral typechecker as well.
It's slow as hell. And every version bump comes with a bunch of new warnings/errors by default that is quite disruptive if you want to keep it quiet. But mostly that it is annoyingly slow.
Might be worth checking out https://github.com/DetachHead/basedpyright
But I’ve had the same issue with too many warnings, mypy is better at understanding Python but even slower.
I guess it’s all relative. One of my code bases still has pylint running with only a couple custom lint rules, that one is slow as hell.
As for version bumping, maybe it’s just a me thing, but I hard fix a version and only update occasionally. Sure each update brings new warnings, but most of them are valid and if you only do it a couple times a year… not that big a deal.
Try ruff instead of pylint. It will open your eyes to how performant software can actually be. All of us that have been working in python and JavaScript for too long have really forgotten.
We use ruff for everything that isn’t a custom rule.
The only thing I found it ‘missing’ was an indentation check (it’s in preview, I don’t turn on preview rules), but I realized it doesn’t matter because we also have a formatter running on everything.
It’s dog slow. I realize we are talking about a python checker written in typescript, so two relatively slow languages combined, but still.
The dependency on node makes it a PITA to set up in non-vs code environments as well.
pyright doesn't do well with inspecting inherited types, at least in our experience. Falls far behind JetBrain's type checker in Pycharm. (This is one reason why pycharm >> zed.)
Are you kidding? Pycharm's pychecker is very incomplete. My whole team moved away from pychram due to the shitty typehint support which barely moved forward. (5+ year old bug, inconsistent behavior across the different features in pycharm indicating to me a fragile mess of heuristics behind the curtains)
I agree that most of the functionality on top of the type checker/engine is better in pycharm though. But a correct implementation is more important
The team behind ruff/uv is working on a similar tool:
https://x.com/charliermarsh/status/1884651482009477368
Have there been any updates on when they think it’ll be released?
At least the alpha milestone is May 12, but I don't think they've publically committed to anything.
In particular for me it would be great to have a better typechecker than Pyright available for Zed (basically why the editor is a nonstarter right now).
An editor is a nonstarter because it doesn't doesn't integrate the right implementation of some particular optional feature for some particular language?
That's a high bar!
I have some sympathies with the parent post as Zed is my daily driver and I mostly write Python. Pyright is horrible to work with, tons of false positives. Parent poster - you might find this useful: https://github.com/zed-industries/zed/discussions/24801
No it is not horrible to work with. Even the “standard” settings are quite lax.
I think our respective comment votes disagree.
Well, I mean why use a suboptimal tool? I don’t have any ideology to follow, my codebase is just mostly Python and heavily typed. Pyright doesn’t support all the features so I stick with VSCode/Pylance.
So what editor is a starter then? As far as I know pyright is the de facto checker for python in most editors. Not sure why it would be worse in zed, seems to work fine for me. Or as well as pyright can work.
This is awesome! Thank you for sharing it!
I'm wondering though, why do existing tools like Mypy and Pyright not scale for Meta? Is Pyre(fly) being used extensively at Meta? In general what sort of issues did Meta run into?
We go into it a bit more here: https://pyrefly.org/en/docs/pyrefly-faq/ The short answer is that the scale of the Python projects we work on is huge. Pyre is the only checker that can scale for our needs today. To bring modern features like type inference and better IDE integration, a full rewrite needed to happen. We are not done just yet, but have been using our massive codebases to test the performance and flexibility of the checker. Thanks for the question!
> pyrefly check --suppress-errors
> INFO 5,240 errors shown, 65,932 errors ignored
Not a single Python type checker had ever worked for me so far.
I don't see the link between that output and your conclusion. Are those results wrong?
> Are those results wrong?
I don't know. Some packages just don't work with type checkers, e.g. Django.
Django has a community stubs project, are you using it?
Does django-stubs work with other type checkers than Pyright?
the stubs can only do so much. There's so much magic and inferred properties that a type checker will never be able to fill in alone
Is it not that he used the "--suppress-errors" flag yet it did not ignore the errors?
It's hard to tell. Other similar options in other commands can mean things like "exit with status 0 even if errors are detected". Without installing pyrefly, I'm not sure what that flag's supposed to do. It's not documented on their site.
It adds a comment to suppress the error above each line in your codebase that causes an error, but it reports the errors before doing that, so you can only see the result after you re-run `pyrefly check`.
Oh! That's nifty. It leaves you with a baseline so you can stop making new mistakes, and gives you something to grep for when you want to fix existing ones.
Thanks for giving this a try! We go over how to use the upgrade tooling (`--suppress-errors`) in the guide here: https://pyrefly.org/en/docs/installation/. This is intended to make upgrading from different versions of type checkers easier. It’s something we use internally and wanted to share with the community along with the checker itself.
We also allow you the ability to suppress whole classes of errors if you want to ignore specific error types and avoid inline ignores: https://pyrefly.org/en/docs/configuration/ A word of caution: this will ignore future errors of this type as well.
First, make sure your project is properly configured and then follow the instructions. `--suppress-errors` will add ignores inline allowing the project to check cleanly. We have thought about a feature that allows you to keep suppressions in a separate file, but we have not put it on the roadmap yet. If this is something important for your workflow, we would like to learn more - please add a feature request on GitHub.
It's literally working? What did you expect?
Maybe it's working but it's not useful.
What would you expect "useful" to be if your codebase is basically incompatible with type checking?
I expected that I would be able to run the check command and it would just work. Upon reading the docs, this tool recommends incremental adoption, and after using `--suppress-errors`, `# pyrefly: ignore` is all over my codebase.
I know I shouldn't get into this thread, but I'm extremely curious: what did you expect should have happened? I mean, literally, what do you mean by "just work", what work did you expect a type checker to perform if not to show you errors?
I expected a reasonable amount of reasonable errors, not 60k+ errors most of which are import-error, even though the code runs fine.
edit: most errors weren't actually import-error, I just misunderstood --search-path
Its the fault of the tool that the codebase had lots of errors?
I would say it's the fault of the ecosystem. Hopefully it will get better as libraries adopt type hints.
That's incremental adoption for adding type hints, not for adopting this particular tool in an already-type-hinted codebase.
…how else would it work?
What does "just work" mean? You are comically obtuse.
Fix your types.
Don't blame the tool for your shoddy work. Either accept your mediocre quality or do something to fix it.
Do you... have type annotations? I think you might be missing the point.
I think the point here is that for something like Python the default behaviour should be an assumed return type of `Any` rather than throwing an error. Maybe that is the case and the GP had configured it otherwise.
No it shouldn't because then you can easily miss places where you should have added a type annotation, and also your lazy colleagues won't bother adding them at all.
The worst of both worlds is colleagues who insert type annotations but they're wrong. And there's no CI pipe to verify the types before check-in.
Working with these things makes me often think that people who want to write high quality software should just use a better language. But, well, real world is real.
> just use a better language
This
Yeah unfortunately Python is crazy popular so a lot of the time you don't get a choice.
If fairness if you set up uv and Ruff and Pyright it's kind of ok. Still have to deal with the general noobness of the ecosystem and the horrific performance, but I've seen worse.