Heavy sync init work (e.g. local ML models) starves stdio initialize/tool calls even with threadpool offload -- Windows GIL contention

[scotthibbs]

Environment

• mcp 1.28.1 (Python SDK, mcp.server.fastmcp.FastMCP)
• stdio transport, Windows 11
• Blocking work: sentence-transformers / torch (local embedding model), but this
  generalizes to any CPU/GIL-bound library load (onnxruntime, local LLM inference, etc.)

Summary

Related to #1839 / #1909, but a distinct failure mode that threadpool offload does not
fix: GIL-bound CPU work (loading a torch model) contends with the anyio stdio
transport's Windows pipe-reader thread for the GIL, regardless of where it runs in-process.
This makes several reasonable-looking designs fail in ways that are easy to ship and hard
to reproduce in a quick manual test.

Measurements

Same model load (sentence-transformers/all-MiniLM-L6-v2), only the location changes:

Where the load runs	initialize	First tool call
Standalone script, no event loop	—	~8s
Synchronously before mcp.run() ("warm on start")	blocks 8–70s	instant
In a background thread, started before mcp.run()	delayed	deadlocks
Lazily on first call, offloaded via anyio.to_thread / threadpool	instant	~70s
In a separate subprocess, talked to over a pipe	~1s	~8–9s

Only the subprocess isolation keeps both connection time and first-call time bounded.
The variance (8s → 70s) is driven by contention between the model-loading thread and
anyio's Windows stdio reader thread — both fighting for the GIL.

Why this matters beyond my case

It's tempting to "fix" a slow first tool call by warming eagerly before mcp.run().
That instead moves the stall onto the initialize handshake — which is worse, because a
client that doesn't get the handshake in time just drops the server with no visible
error. That's a silent failure (tools missing, no exception) that's intermittent
(fine when warm/cached, broken on a cold start), so it's easy to ship and hard to catch
in CI or a quick manual check.

Suggested fix / ask

Not asking for an SDK-level fix necessarily — run_in_threadpool from #1909 is the right
answer for I/O-bound blocking calls. But it'd help other implementers to:

1. Note in the docs/guidance for FastMCP that CPU/GIL-bound initialization (local model
   loads, etc.) should NOT be threadpool-offloaded in-process — it should run in a
   separate subprocess — since threadpool offload does not release the GIL contention
   the way it does for I/O-bound blocking calls.
2. Consider whether initialize should have an explicit, documented deadline/backpressure
   behavior so implementers know exactly how much startup latency is safe.

---

Disclaimer: I'm not a Python/asyncio expert — this diagnosis, the measurements, and
this write-up were produced by Claude Code (Anthropic's coding agent) while it was
building a local memory/retrieval MCP server for me and debugging why it intermittently
failed to connect. I'm filing it because the finding looked substantive and reproducible,
but I likely can't answer deep follow-up questions about the internals myself.