Running out of memory when using a combination of monkey patched asyncio.sleep and MCP Client

[gediminasz]

What happened?

I am making an MCP tool/task which uses await asyncio.sleep when polling some external API. I wanted to add some unit tests using the FastMCP Client and noticed a huge memory usage spikes:

The pytest process needs to be killed otherwise it runs out of memory and my computer freezes.

Example Code

import asyncio
from unittest.mock import patch

from fastmcp import Client, FastMCP

mcp = FastMCP()


@mcp.tool(task=True)
async def sleepy():
    await asyncio.sleep(10)
    return {"success": True}


async def test_sleepy():
    with patch.object(asyncio, "sleep"):
        async with Client(transport=mcp) as client:
            await client.call_tool("sleepy")

Version Information

FastMCP version:                                                                 3.2.3
MCP version:                                                                    1.27.0
Python version:                                                                 3.14.3
Platform:                                Linux-6.8.0-107-generic-x86_64-with-glibc2.39

[jlowin]

Thanks for the repro — the hot loop isn't your tool body, it's fakeredis.

When any component is declared with task=True, FastMCP starts a Docket instance at server lifespan, and the default backend is memory:// (fakeredis). Docket's strikelist monitor polls Redis with XREAD, and fakeredis implements that poll as await asyncio.sleep(0.01) inside a can_read loop. Your patch.object(asyncio, "sleep") replaces sleep globally with an AsyncMock that returns immediately without yielding, so the polling loop becomes unbounded and allocates on every iteration. That's the memory blowup.

Scope the patch to the module under test rather than patching asyncio.sleep globally:

# my_module.py
import asyncio

async def my_tool():
    await asyncio.sleep(10)
    ...

# test
with patch("my_module.asyncio.sleep"):
    ...

That leaves fakeredis's polling loop intact.