ferro-io 0.2.0

Multi-core async runtime for Python, backed by Tokio. Drop-in asyncio replacement.

ferro_io

Multi-core async runtime for Python, backed by Tokio. Ships with ferro_io, a 100% drop-in replacement for asyncio (every public symbol resolves; 313 tests verify it, including real asyncpg / SQLAlchemy / FastAPI integration).

# Mode 1 — library-level
import ferro_io as asyncio
asyncio.run(main())

# Mode 2 — process-wide (third-party libs also benefit)
import ferro_io
ferro_io.install()
import aiofiles            # now uses ferro_io under the hood

Why

CPython's GIL serializes asyncio onto one core. ferro_io moves the scheduler into Tokio with a multi-thread runtime, so:

• IO workloads stay at the theoretical sleep floor (no slower than stdlib).
• CPU workloads that go through ferro_io.AsyncRuntime.map_blocking use tokio::task::spawn_blocking and bypass the GIL entirely.

Benchmarks (M-series Mac, 14 cores)

Best of 5 trials. Full matrix with uvloop in benchmarks/RESULTS.md.

Workload	stdlib asyncio	uvloop	ferro_io	vs stdlib
50 × 50ms IO sleep	51.62 ms	51.57 ms	51.44 ms	~1× (sleep floor)
200 × 20ms IO sleep	22.94 ms	22.00 ms	21.23 ms	~1× (sleep floor)
14 CPU chains × 5M LCG iters	6036 ms	5578 ms	6.37 ms	🔥 947×

IO workloads: ferro_io and uvloop are both pinned at the theoretical sleep floor — neither beats physics. CPU workloads: uvloop (libuv) stays GIL-bound like stdlib because it optimizes the event loop, not CPU parallelism. ferro_io's map_blocking routes through tokio::task::spawn_blocking, which releases the GIL and saturates all worker threads.

Coverage

• Symbol-level: 119/119 asyncio public symbols resolve through ferro_io.
• Real-world programs verified: TaskGroup, subprocess pipelines, TCP client/server with streams, to_thread, Queue producer/consumer, gather with return_exceptions, wait_for, timeout, Runner with contextvars.
• Third-party library smoke test: aiofiles works under ferro_io.install().
• Heavyweight libraries verified against real services under ferro_io.install(): asyncpg (Cython records, prepared statements, transactions, pools), SQLAlchemy async (greenlet sync→async bridge over asyncpg), FastAPI / Starlette (anyio, contextvars middleware, to_thread). See tests/test_heavyweights.py.
• 307 unit tests + 6 heavyweight integration tests, 0 skipped, 0 failed.

Compatibility matrix

	stdlib	uvloop	trio	ferro_io
Drop-in asyncio replacement	—	✅	❌	✅
asyncio symbol coverage	100%	~98%	0%	100% (119/119)
asyncio.TaskGroup (3.11+)	✅	✅	n/a (nurseries)	✅
Multi-core CPU workloads	❌ GIL	❌ GIL	❌ GIL	✅ (via spawn_blocking)
Windows support	✅	❌	✅	✅
httpx / aiohttp / websockets	✅	✅	partial	✅ (tested)
asyncpg (Cython records)	✅	✅	❌	✅ (tested — see test_heavyweights.py)
SQLAlchemy async (greenlets)	✅	✅	❌	✅ (tested)
FastAPI / Starlette	✅	✅	partial	✅ (tested)

All "tested" cells are backed by tests/test_heavyweights.py, which runs in the integration CI job against a real Postgres service container and also executes a matching stdlib-control run so a test-script bug can't masquerade as a ferro_io incompatibility.

Heavyweight compat tests run in the integration CI job against a real Postgres service container. To run them locally:

docker compose -f tests/docker-compose.yml up -d
FERRO_IO_INTEGRATION=1 pytest tests/test_heavyweights.py -v

Benchmark matrix vs uvloop

Generate the full matrix (runs each column in a subprocess so uvloop.install() can't contaminate the others):

pip install uvloop     # optional
python benchmarks/bench_matrix.py    # writes benchmarks/RESULTS.md

See benchmarks/RESULTS.md for the latest numbers. TL;DR: uvloop and ferro_io both sit at the IO sleep floor (within noise of each other). On workload C, ferro_io's spawn_blocking path bypasses the GIL while uvloop — like stdlib — remains serialized.

Caveat for process-wide mode

Code that does from asyncio import sleep at module top-level captures the stdlib reference at import time, before ferro_io.install() can run. To work around it, call ferro_io.install() as the very first statement in your entry point — before any third-party import.

Build

python -m venv .venv && source .venv/bin/activate
pip install maturin pytest pytest-asyncio
maturin develop --release
pytest tests/ -v
python benchmarks/bench.py

Built with PyO3 0.28 + pyo3-async-runtimes 0.28 + Tokio + maturin. ABI3 wheels cover Python 3.9+.