Calling Rust from Python/PyPy using CFFI (C Foreign Function Interface)

rust-python-cffi.md

This is a small demo of how to create a library in Rust and call it from Python (both CPython and PyPy) using the CFFI instead of ctypes.

Based on http://harkablog.com/calling-rust-from-c-and-python.html (dead) which used ctypes

CFFI is nice because:

• Reads C declarations (parses headers)
• Works in both CPython and PyPy (included with PyPy)
• Lower call overhead than ctypes

---

Install Rust

from either of:

• brew install rust
• Multirust

I recommend installing Rust via multirust so that you can compile projects that use unstable features (core) as well as testing for Rust 1.0,1.1, etc backwards compatibility.

The per-directory override mechanism is especially nice.

Build library

rustc treble.rs

This will create a libtreble.dylib

Run the Python client

python test.py 

<cffi.api.FFILibrary_./libtreble.dylib object at 0x1089d5490>
math from rust! 30

test.py

try:
    from cffi import FFI
except ImportError:
    print "pip install cffi, included with PyPy"

ffi = FFI()
lib = ffi.dlopen("./libtreble.dylib")

print lib
# <cffi.api.FFILibrary_./libtreble.dylib object at 0x107f440d0>

ffi.cdef('int treble(int);')

print "math from rust!", lib.treble(10)

treble.rs

#![crate_type = "dylib"]
// compile: rustc treble.rs

#[no_mangle]
pub extern fn treble(value: i32) -> i32 {
    value * 3
}

@devdave

I haven't touched this in years, but there is a repo over here https://github.com/seanjensengrey/rust-jna-example that expands on this for Java.

This repo shows how to expose a Levenshtein library using three methods

• Py03 with Python
• Cffi with Python
• Java JNA with Java

This example handles strings, which depending on the ratio of work to marshaling can solve many problems. After all, nearly every service call already has heavy string based overheads.

I like that this is basically "rustc and go" so I can try pulling out a single, bottle-necked function. If I had to combine large Python and large Rust projects, then PyO3 is 100% the way I would go. For porting one function, instead of putting it up as a FaaS/Lambda/micro-service, I really like this use of CFFI.

I think now I would compile the Levenshtein code to Wasm and load it into Python via Wasmtime. @KevinWhalen totally with you on safe inprocess servers. Composes like a web service, fast like a function call.