simloom
simloom swaps asyncio's event loop for a deterministic one, so real async code runs under a seeded scheduler, a virtual clock and a fault-injecting in-memory network. Any failure replays exactly from one seed, and shrinks to the smallest schedule that still breaks.
$ pip install simloomPython 3.12 · asyncio · pytest plugin · PyPI · uv
0.5.0 on PyPI · Python 3.12+ · Apache-2.0
- 10,000
- seeds in the CI determinism job
- 377
- test functions
- 0
- runtime dependencies
Why asyncio fits
Every interleaving decision in asyncio happens at an await, under a replaceable event loop. Libraries like aiohttp and redis-py bottom out in loop primitives, so they run inside the simulation unchanged — no forked interpreter, no hypervisor.
CI runs an unmodified aiohttp HTTPS server and client over an in-sim TLS handshake, and redis.asyncio against an in-sim Redis.
One tape, one universe
All nondeterminism comes from a single choice tape: which ready callback runs next, whether a packet drops, how long a link delays. Each decision is a bounded integer draw. The idea is borrowed from Hypothesis's choice sequence, and there is deliberately no second source of randomness.
The loop only consumes a draw when more than one callback is ready. When nothing is ready, the clock jumps straight to the next timer — a simulated hour runs in milliseconds.
seed
│
▼
┌──────────────┐ draws ┌──────────────────────────────┐
│ choice tape │◀─────▶│ SimLoop │
│ │ │ callback pick · virtual time │
└──────┬───────┘ │ SimNetwork · fault injection │
│ └──────────────┬───────────────┘
│ │ unmodified asyncio code
│ recorded tape ▼
│ oracles: always · eventually ·
│ serializable · linearizable
▼ │ failure
replay() ◀──── shrink ◀────────────┘ seed
│
▼
┌──────────────┐ draws ┌──────────────────────────────┐
│ choice tape │◀─────▶│ SimLoop │
│ │ │ callback pick · virtual time │
└──────┬───────┘ │ SimNetwork · fault injection │
│ └──────────────┬───────────────┘
│ │ unmodified asyncio code
│ recorded tape ▼
│ oracles: always · eventually ·
│ serializable · linearizable
▼ │ failure
replay() ◀──── shrink ◀────────────┘Search and checking
Search can be a random walk, auto-tuned PCT, or systematic mode — a stateless model checker that enumerates delay-bounded schedules, so a pass is a bounded proof rather than a lucky run.
Oracles go beyond crashes: always/eventually monitors, an Elle-style serializability checker, and a Wing–Gong linearizability checker with Lowe's memoisation that catches stale reads serializability would allow.
Shrinking, and a real CPython bug
The shrinker minimises in shortlex order and re-records every accepted candidate. For schedules it minimises the number of non-FIFO picks — the number a reader actually cares about when reading a failing trace.
examples/bpo42130.py vendors the pre-3.12 asyncio.wait_for and recovers bpo-42130, the race where a cancellation was silently swallowed. The modern stdlib survives the same exploration. CI replays 10,000 seeds twice each and requires the event-log hashes to match.