Skip to content

Async generator entrypoint keeps running after the client disconnects (follow-up to #482) #752

Open

Description

@colamong

Describe the bug

When an async def generator entrypoint is streamed and the client disconnects, the handler is never closed. It keeps running to the end, and if more than 100 events remain, the producer task stays parked in _offer() forever (polling every 10 ms) for the life of the process.

#563 added cleanup for this in the finally of _async_gen_to_sync_gen (stop.set() + task.cancel()), but that finally never runs. Starlette's iterate_in_threadpool never calls close() on the sync generator it iterates, so no GeneratorExit reaches it. The unit test for #563 calls sync_gen.close() by hand (test_app.py, _abandon_after), and the integration test only checks that /ping and a fresh session still work, so this was not caught.

To Reproduce

bedrock-agentcore 1.27.0 (main @ 8467c9a), starlette 1.10.0, uvicorn 0.54.0, Python 3.12, macOS.

# server.py
import asyncio, time
from bedrock_agentcore.runtime import BedrockAgentCoreApp

app = BedrockAgentCoreApp()

@app.entrypoint
async def invoke(payload):
    try:
        for i in range(12):
            await asyncio.sleep(1)
            print(f"{time.time():.1f} step {i + 1}/12", flush=True)
            yield {"step": i + 1}
        print("finished all steps", flush=True)
    except BaseException as e:
        print(f"{time.time():.1f} closed: {type(e).__name__}", flush=True)
        raise

app.run(port=8080)
# client.py: read for 3 seconds, then disconnect
import time, httpx
start = time.time()
with httpx.Client(timeout=30) as c, c.stream("POST", "http://127.0.0.1:8080/invocations", json={}) as r:
    for _ in r.iter_raw():
        if time.time() - start >= 3:
            break
print(f"{time.time():.1f} client disconnected")

The client disconnects after step 3. The server prints steps 4 to 12 and then finished all steps.

With range(500) and asyncio.sleep(0.01), the server stops around step 103/500 and never prints anything again. That is 100 queued events plus a few in flight, with the producer parked in _offer().

Expected behavior

The handler is closed shortly after the disconnect, as the _async_gen_to_sync_gen docstring describes.

Possible fix

Iterate the sync generator ourselves and close it in finally:

-                return StreamingResponse(self._sync_stream_with_error_handling(result), media_type="text/event-stream")
+                return StreamingResponse(self._iterate_sync_stream(result), media_type="text/event-stream")
...
+    async def _iterate_sync_stream(self, generator):
+        stream = self._sync_stream_with_error_handling(generator)
+        try:
+            async for chunk in iterate_in_threadpool(stream):
+                yield chunk
+        finally:
+            stream.close()
+            generator.close()

With this change, the repro above logs closed: CancelledError at step 4, and the 500-step case stops at step 2. I added an integration test, test_handler_is_closed_after_disconnect, to test_streaming_disconnect_integration.py. It fails on main and passes with the change (5/5 runs). test_app.py passes (148).

Integration test (added to tests/integration/runtime/test_streaming_disconnect_integration.py)

The live_server handler records how it exits ("completed" or "closed") per sessionId in a module-level _handler_exits dict, cleared by the fixture.

def test_handler_is_closed_after_disconnect(live_server):
    """A mid-stream client disconnect closes the handler instead of leaving its producer parked."""
    base = live_server

    with httpx.Client(timeout=30.0) as c:
        with c.stream("POST", f"{base}/invocations", json={"sessionId": "A", "chunks": 5000}) as r:
            for i, _ in enumerate(r.iter_lines()):
                if i >= 3:
                    break

    deadline = time.time() + 5
    while "A" not in _handler_exits and time.time() < deadline:
        time.sleep(0.1)
    assert _handler_exits.get("A") == "closed"

Notes

  • Cleanup can be delayed. uvicorn reports ASGI spec 2.3, so Starlette cancels the response through a task group. That cancel waits for the in-flight next() in the thread pool, so cleanup happens when the handler yields its next event. A handler parked in a long call is closed only after that call returns.
  • Sync generator handlers change too. They go through the same path, so they also get closed on disconnect now.
  • Scope. I only checked this locally with uvicorn, not on AgentCore Runtime.

Activity

  1. added
    high-severityOutage, data loss, security issue, crash with no workaround, or regression in a released version
    bugSomething isn't working
    on Nov 21, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workinghigh-severityOutage, data loss, security issue, crash with no workaround, or regression in a released version

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions