simple per second rate limiter

rate_limiter.py

def rate_limit_per_second(rate: int):
    last_request = time.time()
    lock = asyncio.Lock()

    @contextlib.asynccontextmanager
    async def rate_limiter():
        nonlocal last_request, lock
        async with lock:
            now = time.time()
            elapsed = now - last_request
            to_sleep = 1 / rate - elapsed
            if to_sleep > 0.0:
                await asyncio.sleep(to_sleep)
                last_request = max(now + to_sleep, last_request)
            else:
                last_request = max(now, last_request)
        yield None

    return rate_limiter