SSE Hangs After Disconnect #1437

Activity

1. maxisbey commented on Oct 7, 2025

   

   maxisbey

   on Oct 7, 2025

   Contributor

   More actions

   Could you please provide a minimal reproducible example code snippet we can run locally to see the issue?
2. maxisbey

   added

   needs reproneeds additional information to be able to reproduce bugneeds additional information to be able to reproduce bug

   bugSomething isn't workingSomething isn't working

   on Oct 7, 2025
3. cameronsjo commented on Nov 2, 2025

   

   cameronsjo

   on Nov 2, 2025

   Author

   More actions

   Reproduction Attempt & Code Analysis

   I've created a minimal reproduction environment for this issue. While I wasn't able to trigger the bug with current library versions, I can confirm the vulnerable code pattern exists.

   📦 Reproduction Repository

   Repository: https://github.com/cameronsjo/sse-disconnect-repro

   Complete setup with:

   • FastMCP SSE server with ASGI protocol tracking middleware
   • Multiple test clients with various disconnect strategies
   • Full documentation and analysis

   git clone https://github.com/cameronsjo/sse-disconnect-repro
   cd sse-disconnect-repro
   uv sync
   uv run python server.py  # Terminal 1
   uv run python client.py  # Terminal 2

   🔍 Confirming the Vulnerable Code Pattern

   The issue exists in mcp/server/streamable_http.py lines 460-480:

   async def sse_writer():
       try:
           async with sse_stream_writer, request_stream_reader:
               async for event_message in request_stream_reader:  # ← Blocks here
                   event_data = self._create_event_data(event_message)
                   await sse_stream_writer.send(event_data)  # ← Never reached if disconnect early
                   # ...

   The Flow:

   1. Line 497: EventSourceResponse sends http.response.start immediately
   2. Line 465: sse_writer() blocks waiting for events from the MCP server
   3. Client disconnects before any events generated
   4. sse_stream_writer.send() never called → No http.response.body → ASGI violation

   🧪 Test Results

   Tested with:

   • FastMCP: 2.13.0.2
   • MCP SDK: 1.20.0
   • sse-starlette: 3.0.3

   Result: Unable to reproduce - all requests completed correctly.

   Why: MCP server responds in <10ms, and sse-starlette appears to send initial chunks that satisfy ASGI requirements.

   🔧 Proposed Fix

   Even though hard to reproduce, the defensive fix ensures robustness:

   async def sse_writer():
       body_sent = False
       try:
           async with sse_stream_writer, request_stream_reader:
               async for event_message in request_stream_reader:
                   event_data = self._create_event_data(event_message)
                   await sse_stream_writer.send(event_data)
                   body_sent = True
                   if isinstance(event_message.message.root, JSONRPCResponse | JSONRPCError):
                       break
       except Exception:
           logger.exception("Error in SSE writer")
       finally:
           if not body_sent:
               await sse_stream_writer.send({"event": "close", "data": ""})
           logger.debug("Closing SSE writer")
           await self._clean_up_memory_streams(request_id)

   💡 When This Could Trigger

   • Slow MCP processing (>100ms delays)
   • High server load causing task scheduling delays
   • TCP RST packets before buffered data sent
   • Stricter ASGI server implementations

   🎯 Recommendation

   Apply the defensive fix because:

   • Vulnerable pattern exists in the code
   • Edge cases under load could trigger it
   • Simple fix adds robustness
   • Ensures ASGI spec compliance

   Happy to submit a PR with the fix if helpful!
4. cameronsjo commented on Mar 28, 2026

   

   cameronsjo

   on Mar 28, 2026

   Author

   More actions

   Closing — no longer relevant to my workflow.
5. cameronsjo

   closed this as completedon Mar 28, 2026