pyro3 0.6.0

Pyroxide

A lock-free, high-concurrency background task broker for Python, powered by Rust.



Explore the Docs »

API Reference · See Examples · Report Bug · Request Feature

Pyroxide (pyro3) is a lightweight, ultra-high-performance background task broker designed to bridge Python and Rust. It allows CPU-bound or blocking workloads to bypass the Python Global Interpreter Lock (GIL) with minimal memory overhead and zero CPU-sleep polling.

Why Pyroxide?

• 🚀 Bypass the GIL (GIL-Free): Execute CPU-intensive compiled tasks on background OS threads without holding the Python GIL.
• ⚡ Microsecond Latency: Utilizes OS-level signaling (Condvar) rather than CPU-burning thread polling, dispatching and completing tasks in under 25 microseconds.
• 📦 Zero Infrastructure: Runs completely in-process. No Redis, RabbitMQ, or Celery worker daemons to configure or maintain.
• 💾 Zero-Copy Serialization: Pass large byte arrays, memoryviews, or columnar buffers across the C-ABI boundary without copy or pickle overhead.
• 🛠️ On-the-Fly Native Compilers: Write code as Python strings and compile them to dynamic libraries on-the-fly (Rust, C, and Zig supported!).
• 🛡️ Isolated Worker Processes: Opt-in isolated=True to run tasks in separate processes via cross-platform Named Pipes / Domain Sockets. Features bidirectional Zero-Copy Shared Memory (SHM) routing for payloads >= 1MB, and an auto-scaling Scale-to-Zero pool to reclaim memory.
• 🔗 Task Groups & Workflows: Bundle multiple task handles into parallel groups (group) to await or cancel them as a single logical unit.

Pyroxide vs. Alternatives

For a detailed analysis, check out the Library Comparison Guide.

Installation

From PyPI

pip install pyro3

Build Locally

Ensure you have Rust, Python (3.8+), and maturin installed:

git clone https://github.com/emivvvvv/pyroxide.git
cd pyroxide
pip install maturin
maturin develop

Quick Start

1. Offload Python Callables

from pyroxide import task

@task
def calculate_square(x: int) -> int:
    return x * x # Runs in background OS threads

# Submit and get a handle immediately
handle = calculate_square(12)
result = handle.result() # Blocks natively (0% CPU) until complete
print(result) # 144

# Pure Python tasks can fully bypass the GIL with `isolated=True`
@task(isolated=True)
def heavy_computation(x: int) -> int:
    return sum(i * i for i in range(x))

2. Batch Submission & Task Groups

Submit tasks in bulk under a single lock acquisition to avoid thread contention, and manage them concurrently:

from pyroxide import task, group

@task
def calculate_square(x: int) -> int:
    return x * x

payloads = [10, 20, 30, 40]

# 1. Batch submit payloads
handles = calculate_square.batch(payloads)

# 2. Bundle into a parallel TaskGroup
tg = group(handles)
print(tg.status) # "Running"

# 3. Retrieve results (consume=False preserves status metadata)
results = tg.result(consume=False)
print(results)   # [100, 400, 900, 1600]
print(tg.status) # "Completed"

3. Sandboxed WebAssembly (GIL-Free)

Run computations GIL-free in a secure, virtual sandbox without compiling native code:

from pyroxide import register_wasm, wasm_task, load_wasm

# 1. Register WebAssembly bytecode
with open("rot13.wasm", "rb") as f:
    register_wasm("rot13", f.read())

# 2. Execute via decorators
@wasm_task("rot13")
def rot13_cipher(payload: str) -> str:
    pass

print(rot13_cipher("hello").result()) # "uryyb"

# 3. Or load as an Object-Oriented Proxy!
cipher = load_wasm("rot13")
print(cipher.run("hello").result()) # "uryyb"

4. Dynamic Shared Libraries (On-the-Fly Compilation)

Compile and load native code strings on-the-fly. Rust (compile_dylib), C (compile_c), and Zig (compile_zig) are supported:

from pyroxide import compile_dylib, dylib_task, load_dylib

RUST_SRC = """
#[no_mangle]
pub unsafe extern "C" fn pyroxide_plugin_run(ptr: *const u8, len: usize, out_len: *mut usize) -> *mut u8 {
    let input = std::slice::from_raw_parts(ptr, len);
    let s = std::str::from_utf8(input).unwrap_or("");
    let result = s.to_uppercase().into_bytes();
    *out_len = result.len();
    let boxed = result.into_boxed_slice();
    Box::into_raw(boxed) as *mut u8
}

#[no_mangle]
pub unsafe extern "C" fn pyroxide_plugin_free(ptr: *mut u8, len: usize) {
    let _ = Box::from_raw(std::slice::from_raw_parts_mut(ptr, len));
}
"""

# Compile, register and load the Rust library on-the-fly!
compile_dylib("rust_upper", RUST_SRC)

# 1. Execute via decorators
@dylib_task("rust_upper")
def to_upper_rust(payload: str) -> str:
    pass

print(to_upper_rust("hello from rust").result())  # "HELLO FROM RUST"

# 2. Or load as an Object-Oriented Proxy to call any custom C-ABI symbol directly!
rust_upper = load_dylib("rust_upper")
print(rust_upper.pyroxide_plugin_run("hello from rust").result())  # "HELLO FROM RUST"

Dive Deeper (Documentation Book)

Detailed documentation, guides, and implementation examples are available in our Documentation Book:

• Asynchronous Event Loops: Non-blockingly await tasks using await handle.result_async() in FastAPI/asyncio. Read Chapter.
• Isolated Worker Processes: Sandbox tasks in separate OS processes for crash safety and GIL bypass. Read Chapter.
• Batch Submissions: Submit multiple tasks under a single lock acquisition to avoid thread contention. Read Chapter.
• Task Cancellation: Gracefully abort long-running background tasks mid-flight. Read Chapter.
• Traceback Preservation: Capture stack traces on background worker threads and propagate them to the main thread. Read Chapter.
• Memory Footprint & GC: Learn how Slab memory is reclaimed automatically using GC destructors. Read Chapter.

Performance At-a-Glance

We benchmarked Pyroxide against CPython's standard concurrency pools using identical compute payloads (recursive Fibonacci 20 workload) on Apple M1 Pro (8 cores, 16GB RAM):

• Bypassing the Multiprocessing Bottleneck: While Python's ProcessPoolExecutor takes over 2 seconds due to slow process spawning and heavy pickle IPC serialization, Pyroxide's @dylib_task runs native compiled plugins in just 20 milliseconds—offering a 100x speedup with zero-copy shared memory.

Real-World Odoo Enterprise Arrow Ledger Audit Benchmark

To test performance under realistic enterprise data movement workloads, we ran a simulated Odoo Ledger Audit benchmark processing a 9.62 MB Apache Arrow serialized transaction recordset (200,000 journal items) across 10 concurrent requests comparing different concurrency strategies:

• CPython ThreadPoolExecutor (GIL-Locked): 0.3221 s
• Pyroxide Threaded @task (GIL-Locked): 0.3298 s (matches Python's native scheduling overhead perfectly)
• ProcessPoolExecutor (Python, Pickled Pipes): 0.2758 s
• Pyroxide SHM Isolated @task (Zero-Copy SHM): 0.3272 s
• Pyroxide @dylib_task (C-compiled, GIL-Free): 0.0091 s (bypasses GIL entirely)

Key Takeaway: By offloading the audit logic to a dynamically compiled C/Rust plugin running on Pyroxide's background thread pool, we achieve a 35.3x speedup over CPython's standard ThreadPoolExecutor by completely bypassing the GIL.

To run the Odoo simulation suite locally:

python examples/odoo_poc/odoo_complex_simulation.py

To run the comparative and basic benchmark suites locally:

# 1. Run detailed comparative benchmarks against CPython concurrency pools
python examples/benchmarks/benchmark_vs_alternatives.py

# 2. Run basic scheduling latency and asyncio benchmarks
python examples/benchmarks/benchmark.py

Contributing

Contributions are welcome! If you'd like to improve Pyroxide or add support for additional features, feel free to open an issue or submit a pull request on GitHub.

License

Pyroxide is licensed under any of:

• MIT License (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)
• Coffeeware License (LICENSE-COFFEE)

at your option.