I appreciate that you first tried to optimize the original Python code. Idiomatic Python is unfortunately disappointingly slow and not so interesting to compare to.
I often use the rough approximation that Python is 40-50x slower than C. This is what you'll see in the benchmarks.
The truly rough thing about Python though is that that is the speed when the code is being written to a benchmark. It is really, really easy to write Python that is multiples slower than that when not writing to a benchmark and just trying to get work done without hyperoptimizing. I did some testing of Python [1] to back some other commentary I was making that compared the time it took to set an attribute repeatedly on a particular instance of an empty class to the time it took to setting it on a subclass of a subclass of a class that had a property setter that was wrapped by a decorator. The latter was about 4.6 time slower than the direct attribute setting, which was itself already ~100x slower than an attribute setting in a static language.
And it's not like a three-deep nested class with a property wrapped by a decorator is all that absurd in Python or anything. That's a completely normal case, not some absurd example I made up to skew the test.
In practice the 40-50x number is more lower bound than what you can count on. If you are actually using Python's features I think you can easily score another order of magnitude slower without anything jumping out at you as being an obviously bad idea.
> 40-50x number is more lower bound [...] easily score another order of magnitude slower
This is about what I observe. I had a utility based on `scapy`; there were no obviously bad ideas in the python source, but porting the work loop into a cpython extension module yielded a 500x speedup.
I feel like scapy is a bit of a special case though. It's unusually dynamic because it allows kind of arbitrarily crazy things. I feel like it's maily best for prototyping. I've had 1000x speedups (100s of milliseconds to dozens of microseconds) just by removing dynamism. In our case we used scapy to work out the packet shapes we cared about and then just implemented just that subset (still in Python!). But that dynamism really helped with the early stages along with wireshark
A while back I had claude implement something, I don't quite remember what it was, but it chose Python. It was going to take hours. I told it to rewrite it in Rust and it was > 300000x faster. This is without any optimization or prompting particularly about performance, a short one shot lift.
echo "Python sucks, use something else when you can" >> ~/CLAUDE.md
Python was cool in 2005 in academia IT, all the rage in startup 2012. These days...
I've generally taken the position that if you're considering doing micro optimizations in your python code you should just switch to another language instead. The gains from switching language are going to be so much higher than the gains from trying to get the python interpreter to be a little less slow (and empirically, every time I've seen it tried the code has been rewritten in the end anyway).
I mean that mostly in the sense that there is huge variance in idiomatic code. So your optimized C/Rust code might be 100-1000x faster than two idiomatic versions of writing that code
The second part, where it says that Rust is faster than Python, is so obvious than doesn't deserve any further comments. What I found interesting, is that idiomatic Python, with generators and all, can be so much worse than "C-style" Python, if you will pardon that term. It's a pity many kids are taught Python before C...
I really enjoy making Python faster. I feel like the sweet spot for me is proving a concept with the dumbest possible implementation to show that something would work, and then using that as a comparison implementation to prove that later improvements match the results of the dumb, obviously correct implementation.
it's possible! although many of the constraints in this blog were because we wanted to work with ast module in Python. If we were allowed to create our own types, we can do so much better. I think ruff has an even faster walk by those standards.
It seems bandit is using some decent optimizations already, looking at the `@test.checks("Call")` seems like they already captured some easy wins.
The largest win honestly would be using the same ast.walk for multiple rules, which we also did, but not mentioned in the blog.
Just rewrote docutils and myst-md-parser in rust with byte-for-byte equivalent output the other day; as westurner/dsport/src/docutilsrs and sphinxdocrs and so on. Pygmentsrs required fancy-regex because rust regex is linear. rstest and cargo-insta helped.
You’d have to spend a bit of time picking apart your arguments and releasing stuff on early error returns, but otherwise the Python/C API is completely banal and doesn’t really have much to learn. It might be worth the time to poke at it at some point to assure yourself this is the case.
(Or you could switch to C++ and use pybind11, but now you’re just switching from one quite complex and somewhat off-putting language to another really complicated and very ugly one, so the win is less clear.)
I appreciate that you first tried to optimize the original Python code. Idiomatic Python is unfortunately disappointingly slow and not so interesting to compare to.
I often use the rough approximation that Python is 40-50x slower than C. This is what you'll see in the benchmarks.
The truly rough thing about Python though is that that is the speed when the code is being written to a benchmark. It is really, really easy to write Python that is multiples slower than that when not writing to a benchmark and just trying to get work done without hyperoptimizing. I did some testing of Python [1] to back some other commentary I was making that compared the time it took to set an attribute repeatedly on a particular instance of an empty class to the time it took to setting it on a subclass of a subclass of a class that had a property setter that was wrapped by a decorator. The latter was about 4.6 time slower than the direct attribute setting, which was itself already ~100x slower than an attribute setting in a static language.
And it's not like a three-deep nested class with a property wrapped by a decorator is all that absurd in Python or anything. That's a completely normal case, not some absurd example I made up to skew the test.
In practice the 40-50x number is more lower bound than what you can count on. If you are actually using Python's features I think you can easily score another order of magnitude slower without anything jumping out at you as being an obviously bad idea.
[1]: https://jerf.org/iri/post/2024/not_about_python_addendum/
> 40-50x number is more lower bound [...] easily score another order of magnitude slower
This is about what I observe. I had a utility based on `scapy`; there were no obviously bad ideas in the python source, but porting the work loop into a cpython extension module yielded a 500x speedup.
I feel like scapy is a bit of a special case though. It's unusually dynamic because it allows kind of arbitrarily crazy things. I feel like it's maily best for prototyping. I've had 1000x speedups (100s of milliseconds to dozens of microseconds) just by removing dynamism. In our case we used scapy to work out the packet shapes we cared about and then just implemented just that subset (still in Python!). But that dynamism really helped with the early stages along with wireshark
A while back I had claude implement something, I don't quite remember what it was, but it chose Python. It was going to take hours. I told it to rewrite it in Rust and it was > 300000x faster. This is without any optimization or prompting particularly about performance, a short one shot lift.
echo "Python sucks, use something else when you can" >> ~/CLAUDE.md
Python was cool in 2005 in academia IT, all the rage in startup 2012. These days...
I've generally taken the position that if you're considering doing micro optimizations in your python code you should just switch to another language instead. The gains from switching language are going to be so much higher than the gains from trying to get the python interpreter to be a little less slow (and empirically, every time I've seen it tried the code has been rewritten in the end anyway).
> not so interesting to compare to
Absolutely disagree here, something that is considered good practice is very interesting to compare to!
I mean that mostly in the sense that there is huge variance in idiomatic code. So your optimized C/Rust code might be 100-1000x faster than two idiomatic versions of writing that code
It’s amazing how much Python punishes you for modularizing your code
The second part, where it says that Rust is faster than Python, is so obvious than doesn't deserve any further comments. What I found interesting, is that idiomatic Python, with generators and all, can be so much worse than "C-style" Python, if you will pardon that term. It's a pity many kids are taught Python before C...
I still recommend to use idiomatic Python though. If you like C-style Python, you can use Cython and get a performance boost at the same time.
I really enjoy making Python faster. I feel like the sweet spot for me is proving a concept with the dumbest possible implementation to show that something would work, and then using that as a comparison implementation to prove that later improvements match the results of the dumb, obviously correct implementation.
I wonder if these lints could have been expressed as semgrep rules?
Could this ast.sprint ast.walk optimization make libCST or bandit faster? https://news.ycombinator.com/item?id=39111747
libCST: https://github.com/Instagram/LibCST
bandit: https://github.com/PyCQA/bandit
Links to codemod tools; "Baby Steps into Genetic Programming" https://news.ycombinator.com/item?id=43617655
it's possible! although many of the constraints in this blog were because we wanted to work with ast module in Python. If we were allowed to create our own types, we can do so much better. I think ruff has an even faster walk by those standards.
It seems bandit is using some decent optimizations already, looking at the `@test.checks("Call")` seems like they already captured some easy wins.
The largest win honestly would be using the same ast.walk for multiple rules, which we also did, but not mentioned in the blog.
ruff probably does.
Docs for ruff_python_ast: https://docs.rs/littrs-ruff-python-ast/latest/ruff_python_as...
src: ruff/crates/ruff_python_ast: https://github.com/astral-sh/ruff/tree/main/crates/ruff_pyth...
ast.toml, generate.py, Cargo.toml, src/ https://github.com/astral-sh/ruff/blob/main/crates/ruff_pyth...
Just rewrote docutils and myst-md-parser in rust with byte-for-byte equivalent output the other day; as westurner/dsport/src/docutilsrs and sphinxdocrs and so on. Pygmentsrs required fancy-regex because rust regex is linear. rstest and cargo-insta helped.
PyCQA/docformatter https://github.com/PyCQA/docformatter and https://github.com/PyCQA/doc8 would be useful in rust, too. This from the other day: "A benchmark for catching when code doesn't do what its documentation claims"; docs fidelity evals https://news.ycombinator.com/item?id=48530786
AST: Abstract Syntax Tree
FST: Full Syntax Tree
CST: Concrete Syntax Tree
Comment preservation is a feature
[flagged]
only because I'm too lazy to learn how to write C with Python, if anything Rust wasn't helpful with all of those unsafes
You’d have to spend a bit of time picking apart your arguments and releasing stuff on early error returns, but otherwise the Python/C API is completely banal and doesn’t really have much to learn. It might be worth the time to poke at it at some point to assure yourself this is the case.
(Or you could switch to C++ and use pybind11, but now you’re just switching from one quite complex and somewhat off-putting language to another really complicated and very ugly one, so the win is less clear.)
"AI reflex-app builder" slop alert