Speed Up Your Python Using Rust

What Is Rust?

Rust is a systems programming language that runs blazingly fast, prevents segfaults, and guarantees thread safety.

Featuring:

• Zero-cost abstractions
• Move semantics
• Guaranteed memory safety
• Threads without data races
• Trait-based generics
• Pattern matching
• Type inference
• Minimal runtime
• Efficient C bindings

This description is taken from rust-lang.org.

Why Does It Matter for a Python Developer?

I heard a better description of Rust from Elias (a member of the Rust Brazil Telegram Group):

Rust is a language that allows you to build high level abstractions, but without giving up low-level control – that is, control of how data is represented in memory, control of which threading model you want to use etc.
Rust is a language that can usually detect, during compilation, the worst parallelism and memory management errors (such as accessing data on different threads without synchronization, or using data after they have been deallocated), but gives you a hatch escape in the case you really know what you’re doing.
Rust is a language that, because it has no runtime, can be used to integrate with any runtime; you can write a native extension in Rust that is called by a program node.js, or by a python program, or by a program in ruby, lua etc. and, however, you can script a program in Rust using these languages. — Elias Gabriel Amaral da Silva

There are a bunch of Rust packages out there to help you extending Python with Rust. I can mention Milksnake created by Armin Ronacher (the creator of Flask) and also PyO3, the Rust bindings for Python interpreter (see a complete reference list at the bottom of this article).

Let’s see it in action.

For this post, I am going to use Rust Cpython, the only one I have tested; it is compatible with the stable version of Rust and I found it straightforward to use.

NOTE: PyO3 is a fork of rust-cpython and comes with many improvements, but works only with the nightly version of Rust, so I preferred to use the stable for this post. Anyway, the examples here must work also with PyO3.

Pros: It is easy to write Rust functions and import from Python and as you will see by the benchmarks it worth in terms of performance.

Cons: The distribution of your project/lib/framework will demand the Rust module to be compiled on the target system because of variation of environment and architecture. There will be a compiling stage which you don’t have when installing Pure Python libraries; you can make it easier using rust-setuptools or MilkSnake to embed binary data in Python Wheels.

Python Is Sometimes Slow

Yes, Python is known for being “slow” in some cases and the good news is that this doesn’t really matter depending on your project goals and priorities. For most projects, this detail will not be very important.

However, you may face the rare case where a single function or module is taking too much time and is detected as the bottleneck of your project performance, often happens with string parsing and image processing.

Example

Let’s say you have a Python function which does a string processing, take the following easy example of counting pairs of repeated chars, but have in mind that this example can be reproduced with other string processing functions or any other generally slow process in Python.

# How many subsequent-repeated group of chars are in the given string? 
abCCdeFFghiJJklmnopqRRstuVVxyZZ... {millions of chars here}
  1   2    3        4    5   6

Python is slow for doing large string processing, so you can use pytest-benchmark to compare a Pure Python (with Iterator Zipping) function versus a Regexp implementation.

# Using a Python3.6 environment
$ pip3 install pytest pytest-benchmark

Then, write a new Python program called doubles.py

import re
import string
import random

# Python ZIP version
def count_doubles(val):
    total = 0
    for c1, c2 in zip(val, val[1:]):
        if c1 == c2:
            total += 1
    return total


# Python REGEXP version
double_re = re.compile(r'(?=(.)\1)')

def count_doubles_regex(val):
    return len(double_re.findall(val))


# Benchmark it
# generate 1M of random letters to test it
val = ''.join(random.choice(string.ascii_letters) for i in range(1000000))

def test_pure_python(benchmark):
    benchmark(count_doubles, val)

def test_regex(benchmark):
    benchmark(count_doubles_regex, val)

Run pytest to compare:

$ pytest doubles.py                                                                                                           
================================================================================= test session starts ==================================================================================
platform linux -- Python 3.6.0, pytest-3.2.3, py-1.4.34, pluggy-0.4.0
benchmark: 3.1.1 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /Projects/rustpy, inifile:
plugins: benchmark-3.1.1
collected 2 items

doubles.py ..


--------------------------------------------------------------------------------- benchmark: 2 tests --------------------------------------------------------------------------------
Name (time in ms)         Min                Max               Mean            StdDev             Median               IQR            Outliers      OPS            Rounds  Iterations
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_regex            24.6824 (1.0)      32.3960 (1.0)      27.0167 (1.0)      1.8610 (1.0)      27.2148 (1.0)      2.9345 (4.55)         16;1  37.0141 (1.0)          36           1
test_pure_python      51.4964 (2.09)     62.5680 (1.93)     52.8334 (1.96)     2.3630 (1.27)     52.2846 (1.92)     0.6444 (1.0)           1;2  18.9274 (0.51)         20           1
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
=============================================================================== 2 passed in 4.10 seconds ===============================================================================

Let’s take the Median for comparison:

• Regexp – 27.2148 <– less is better
• Python Zip – 52.2846

Extending Python with Rust

Create a New Crate

"Crate" is what we call a Rust package.

Have Rust installed ( the recommended way is https://www.rustup.rs/). I used rustc 1.21.0.

In the same folder, run

cargo new pyext-myrustlib

It creates a new Rust project in that same folder called pyext-myrustlib containing the Cargo.toml (cargo is the Rust package manager) and also a src/lib.rs (where we write our library implementation).

Edit Cargo.toml

It will use the rust-cpython crate as dependency and tell cargo to generate a dylib to be imported from Python.

[package]
name = "pyext-myrustlib"
version = "0.1.0"
authors = ["Bruno Rocha <rochacbruno@gmail.com>"]

[lib]
name = "myrustlib"
crate-type = ["dylib"]

[dependencies.cpython]
version = "0.1"
features = ["extension-module"]

Edit src/lib.rs

What we need to do:

1. Import all macros from cpython crate.
2. Take Python and PyResult types from CPython into our lib scope.
3. Write the count_doubles function implementation in Rust, note that this is very similar to the Pure Python version except for:
   • It takes a Python as first argument, which is a reference to the Python Interpreter and allows Rust to use the Python GIL.
   • Receives a &str typed val as reference.
   • Returns a PyResult which is a type that allows the rise of Python exceptions.
   • Returns an PyResult object in Ok(total) (Result is an enum type that represents either success (Ok) or failure (Err)) and as our function is expected to return a PyResult the compiler will take care of wrapping our Ok on that type. (note that our PyResult expects a u64 as return value).
4. Using py_module_initializer! macro, we register new attributes to the lib, including the __doc__ and also we add the count_doubles attribute referencing our Rust implementation of the function.
   • Attention to the names libmyrustlib, initlibmyrustlib, and PyInit.
   • We also use the try! macro, which is the equivalent to Python’stry.. except.
   • Return Ok(()) – The () is an empty result tuple, the equivalent of None in Python.

#[macro_use]
extern crate cpython;

use cpython::{Python, PyResult};

fn count_doubles(_py: Python, val: &str) -> PyResult<u64> {
    let mut total = 0u64;

    for (c1, c2) in val.chars().zip(val.chars().skip(1)) {
        if c1 == c2 {
            total += 1;
        }
    }

    Ok(total)
}

py_module_initializer!(libmyrustlib, initlibmyrustlib, PyInit_myrustlib, |py, m | {
    try!(m.add(py, "__doc__", "This module is implemented in Rust"));
    try!(m.add(py, "count_doubles", py_fn!(py, count_doubles(val: &str))));
    Ok(())
});

Now Let’s Build It With Cargo

$ cargo build --release
    Finished release [optimized] target(s) in 0.0 secs

$ ls -la target/release/libmyrustlib*
target/release/libmyrustlib.d
target/release/libmyrustlib.so*  <-- Our dylib is here

Now let’s copy the generated .so lib to the same folder where our doubles.py is located.

NOTE: on Fedora, you must get a .so in other system you may get a .dylib and you can rename it changing extension to .so.

$ cd ..
$ ls
doubles.py pyext-myrustlib/

$ cp pyext-myrustlib/target/release/libmyrustlib.so myrustlib.so

$ ls
doubles.py myrustlib.so pyext-myrustlib/

Having the myrustlib.so in the same folder or added to your Python path allows it to be directly imported, transparently as it was a Python module. 

Importing From Python and Comparing the Results

Edit your doubles.py now importing our Rust implemented version and adding a benchmark for it.

import re
import string
import random
import myrustlib   #  <-- Import the Rust implemented module (myrustlib.so)


def count_doubles(val):
    """Count repeated pair of chars ins a string"""
    total = 0
    for c1, c2 in zip(val, val[1:]):
        if c1 == c2:
            total += 1
    return total


double_re = re.compile(r'(?=(.)\1)')


def count_doubles_regex(val):
    return len(double_re.findall(val))


val = ''.join(random.choice(string.ascii_letters) for i in range(1000000))


def test_pure_python(benchmark):
    benchmark(count_doubles, val)


def test_regex(benchmark):
    benchmark(count_doubles_regex, val)


def test_rust(benchmark):   #  <-- Benchmark the Rust version
    benchmark(myrustlib.count_doubles, val)

Benchmark:

$ pytest doubles.py
================================================================================= test session starts ==================================================================================
platform linux -- Python 3.6.0, pytest-3.2.3, py-1.4.34, pluggy-0.4.0
benchmark: 3.1.1 (defaults: timer=time.perf_counter disable_gc=False min_rounds=5 min_time=0.000005 max_time=1.0 calibration_precision=10 warmup=False warmup_iterations=100000)
rootdir: /Projects/rustpy, inifile:
plugins: benchmark-3.1.1
collected 3 items

doubles_rust.py ...


--------------------------------------------------------------------------------- benchmark: 3 tests ---------------------------------------------------------------------------------
Name (time in ms)         Min                Max               Mean            StdDev             Median               IQR            Outliers       OPS            Rounds  Iterations
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
test_rust              2.5555 (1.0)       2.9296 (1.0)       2.6085 (1.0)      0.0521 (1.0)       2.5935 (1.0)      0.0456 (1.0)         53;23  383.3661 (1.0)         382           1
test_regex            25.6049 (10.02)    27.2190 (9.29)     25.8876 (9.92)     0.3543 (6.80)     25.7664 (9.93)     0.3020 (6.63)          4;3   38.6285 (0.10)         40           1
test_pure_python      52.9428 (20.72)    56.3666 (19.24)    53.9732 (20.69)    0.9248 (17.75)    53.6220 (20.68)    1.4899 (32.70)         6;0   18.5277 (0.05)         20           1
--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Legend:
  Outliers: 1 Standard Deviation from Mean; 1.5 IQR (InterQuartile Range) from 1st Quartile and 3rd Quartile.
  OPS: Operations Per Second, computed as 1 / Mean
=============================================================================== 3 passed in 5.19 seconds ===============================================================================

Let’s take the Median for comparison:

• Rust – 2.5935 <– less is better
• Regexp – 25.7664
• Python Zip – 53.6220

Rust implementation can be 10x faster than Python Regex and 21x faster than Pure Python Version. Interesting that the Regex version is only 2x faster than Pure Python.

NOTE: Those numbers makes sense only for this particular scenario; for other cases that comparison may be different.

Conclusion

Rust may not be yet the general purpose language of choice by its level of complexity and may not be the better choice yet to write common simple applications such as web sites and test automation scripts. However, for specific parts of the project where Python is known to be the bottleneck and your natural choice would be implementing a C/C++ extension, writing this extension in Rust seems easy and better to maintain.

There are still many improvements to come in Rust and lots of others crates to offer Python <--> Rust integration. Even if you are not including the language in your tool belt right now, it is really worth to keep an eye open to the future!

References

The code snippets for the examples showed here are available in this GitHub repo.

The examples in this publication are inspired by this Extending Python with Rust talk by Samuel Cormier-Iijima in Pycon Canada video, and also by My Python is a little Rust-y by Dan Callahan in Pycon Montreal (video here).

Other references:

• https://github.com/mitsuhiko/snaek
• https://github.com/PyO3/pyo3
• https://pypi.python.org/pypi/setuptools-rust
• https://github.com/mckaymatt/cookiecutter-pypackage-rust-cross-platform-publish
• http://jakegoulding.com/rust-ffi-omnibus/
• https://github.com/urschrei/polylabel-rs/blob/master/src/ffi.rs
• https://bheisler.github.io/post/calling-rust-in-python/
• https://github.com/saethlin/rust-lather

Join the Community:

Join Rust community, you can find group links here.

If you speak Portuguese, I recommend you join https://t.me/rustlangbr and on .