python-asyncio-pitfalls

name	python-asyncio-pitfalls
description	Use when debugging asyncio event-loop hangs, blocking calls inside coroutines, asyncio.gather vs TaskGroup choice, cancellation handling, structured concurrency, mixing sync and async, run_in_executor decisions, asyncio task lifetimes, "Task was destroyed but it is pending" warnings, or async context manager bugs. Triggers: event loop blocked by sync I/O, ConnectionResetError on cancellation, asyncio in libraries that also offer sync API, FastAPI/aiohttp performance regressions, queue.Queue used in async code. NOT for trio/anyio (different paradigm), threading without asyncio, or Python 2 sync patterns.
category	AI & Machine Learning
tags	["python","asyncio","concurrency","performance","taskgroup","structured-concurrency"]

Python Asyncio Pitfalls

Asyncio is cooperative — one blocking call freezes the entire event loop. Most "asyncio is slow" stories are actually "we accidentally called a synchronous function inside a coroutine." This catalogs the traps.

When to use

• Performance debug: a single slow request blocks all others.
• RuntimeError: This event loop is already running or "Task was destroyed but it is pending".
• Choosing between asyncio.gather, asyncio.TaskGroup (3.11+), and asyncio.wait.
• Cancellation isn't propagating; child tasks keep running.
• Mixing a sync library with async code without freezing the loop.

Core capabilities

Detect blocking calls in the event loop

import asyncio

async def main():
    loop = asyncio.get_running_loop()
    loop.set_debug(True)         # warns when a coroutine takes too long
    # PYTHONASYNCIODEBUG=1 also enables debug mode at the env var level

Debug mode logs:

Executing <Task ... took 0.250 seconds

For production, instrument with loop.slow_callback_duration = 0.1. Anything over 100ms means a blocking call slipped in.

TaskGroup — structured concurrency (Python 3.11+)

async def fetch_all(urls: list[str]) -> list[str]:
    async with asyncio.TaskGroup() as tg:
        tasks = [tg.create_task(fetch(u)) for u in urls]
    return [t.result() for t in tasks]   # exceptions raised here

If any task raises, all siblings are cancelled and the exception (or ExceptionGroup) propagates out. This is the safe default.

asyncio.gather semantics:

results = await asyncio.gather(fetch(a), fetch(b), return_exceptions=True)
# Without return_exceptions=True, the first exception cancels nothing — siblings keep running.

Use gather only when you want the older fire-and-collect-errors semantics; TaskGroup for everything else.

Mixing sync code

# WRONG — blocks the loop.
async def slow():
    return time.sleep(2)         # NOT awaitable; runs sync, blocks loop

# RIGHT — offload to a thread.
async def slow():
    return await asyncio.to_thread(time.sleep, 2)

asyncio.to_thread (3.9+) runs the function in a thread from the default executor. For CPU-bound work, use a ProcessPoolExecutor and loop.run_in_executor:

import concurrent.futures
loop = asyncio.get_running_loop()
with concurrent.futures.ProcessPoolExecutor() as pool:
    result = await loop.run_in_executor(pool, expensive_pure_function, arg)

Cancellation that actually works

async def with_cleanup():
    try:
        await long_operation()
    except asyncio.CancelledError:
        await cleanup()
        raise                    # re-raise so the caller knows you were cancelled
    return result

Swallowing CancelledError is a common cause of "task was destroyed but it is pending" warnings. Always re-raise unless you're explicitly catching to ignore.

Context propagation across tasks

import contextvars

request_id = contextvars.ContextVar('request_id')

async def handler(rid: str):
    request_id.set(rid)
    # Spawned tasks inherit ContextVar values automatically.
    asyncio.create_task(log_async())

Use ContextVars for request-scoped data instead of thread-locals; threading.local doesn't work right in asyncio.

Async iterators and generators

async def stream_chunks(reader: asyncio.StreamReader):
    while not reader.at_eof():
        chunk = await reader.read(4096)
        if not chunk:
            break
        yield chunk

async for chunk in stream_chunks(reader):
    process(chunk)

async for is the right way to consume async iterators; pulling them with __anext__ directly is rarely correct.

Timeouts

# 3.11+ — preferred
async with asyncio.timeout(5):
    result = await long_call()

# Older — wait_for
result = await asyncio.wait_for(long_call(), timeout=5)

asyncio.timeout integrates with TaskGroup; wait_for doesn't and can cause a double-cancel surprise.

Queues

asyncio.Queue is the right choice; queue.Queue is sync and would block.

queue: asyncio.Queue[Job] = asyncio.Queue(maxsize=100)

async def producer():
    while item := await source():
        await queue.put(item)

async def consumer():
    while True:
        job = await queue.get()
        try:
            await process(job)
        finally:
            queue.task_done()

await queue.join()   # wait until all submitted jobs marked done

maxsize provides backpressure — without it, a fast producer can OOM you.

Async context managers and locks

lock = asyncio.Lock()

async with lock:
    await critical_section()

Don't share an asyncio.Lock across event loops; each loop has its own.

FastAPI / aiohttp specifics

• Use async def route handlers when you have async I/O. Sync handlers run in a threadpool — fine, but you pay the thread overhead.
• Connection pools (httpx.AsyncClient, asyncpg pool) belong at app startup, shared across requests. Don't construct per-request.
• Background tasks: BackgroundTasks (FastAPI) for fire-and-forget; asyncio.create_task if you need to track.

Anti-patterns

Calling sync I/O inside a coroutine

Symptom: Throughput craters under load; one slow request blocks everything. Diagnosis: requests.get, time.sleep, psycopg2.execute — all sync, all block the loop. Fix: Switch to async libraries (httpx.AsyncClient, asyncio.sleep, asyncpg). For sync libraries you can't replace, asyncio.to_thread.

Forgetting to await a coroutine

Symptom: coroutine 'fn' was never awaited warning. No error, no return value. Diagnosis: result = fn() instead of result = await fn(). Fix: Add the await. Linters (ruff, pylint) flag this; turn the rule on.

asyncio.gather without return_exceptions=True when you want to wait for all

Symptom: Partial work; some tasks ran, others mysteriously didn't. Diagnosis: First exception propagates and the siblings are NOT cancelled — they keep running detached. Fix: Use TaskGroup (cancels siblings cleanly) or gather(*coros, return_exceptions=True) followed by manual handling.

Swallowing CancelledError

Symptom: "Task was destroyed but it is pending" warning at shutdown. Diagnosis: A task caught CancelledError without re-raising; the runtime never finished cleanup. Fix: Always re-raise after cleanup. Use try/finally for cleanup that must run regardless.

Sharing an event loop across threads

Symptom: RuntimeError: This event loop is already running or sporadic data corruption. Diagnosis: Calling loop.run_until_complete from a thread other than the loop's owner. Fix: Use asyncio.run_coroutine_threadsafe(coro, loop) from non-loop threads.

Unbounded asyncio.Queue

Symptom: OOM under burst load. Diagnosis: Producer faster than consumer; queue grows unbounded. Fix: Set maxsize on the queue. Producer awaits put and naturally blocks.

Quality gates

• No sync I/O calls inside async code paths (linted via ruff ASYNC1* rules).
• TaskGroup used for fan-out where any failure should cancel siblings.
• asyncio.timeout(...) (or wait_for) wraps every external call.
• Connection pools constructed at startup; shared per process.
• CancelledError always re-raised after cleanup.
• asyncio.Queue(maxsize=…) has explicit backpressure.
• loop.set_debug(True) enabled in dev; slow_callback_duration alerts in prod.
• ContextVars used for request-scoped state.

NOT for

• trio / anyio — different structured-concurrency primitives; different mental model.
• threading without asyncio — multiprocessing/threading have different deadlock patterns.
• Python 2 sync patterns — different language era.
• gevent / eventlet — monkey-patching greenlets, separate ecosystem.