What version of gRPC and what language are you using?
- grpcio: 1.76.0
- grpcio-tools: 1.76.0
- Language: Python
What operating system (Linux, Windows,...) and version?
- macOS 26 (Darwin 25.2.0, arm64) —
errno 35 (EAGAIN/EWOULDBLOCK)
- Linux (Ubuntu 22.04, x86_64 c8i.2xlarge) —
errno 11
What runtime / compiler are you using (e.g. python version or version of gcc)
- Python 3.11.13
- uvloop 0.22.1 (also reproduces on the default asyncio
_UnixSelectorEventLoop)
- Granian 2.2.5 ASGI server
What did you do?
Used grpc.aio as the inference-client transport in a production
bid-agent service. Topology per process:
- Granian worker loop (uvloop) handles incoming HTTP
/quote
requests. Each request fires 1-2 unary Predict RPCs plus a
Predict for win-price prediction via a grpc.aio channel pool.
- Background asyncio loop (default
_UnixSelectorEventLoop in a
separate thread) handles model loading, which also makes
LoadModel RPCs through the same channel-manager singleton.
Channel options:
grpc.aio.insecure_channel(
endpoint,
options=[
('grpc.keepalive_time_ms', 10000),
('grpc.keepalive_timeout_ms', 5000),
('grpc.keepalive_permit_without_calls', True),
('grpc.http2.max_pings_without_data', 0),
('grpc.max_receive_message_length', 50 * 1024 * 1024),
('grpc.max_send_message_length', 50 * 1024 * 1024),
],
)What did you expect to see?
grpc.aio to handle FD-readiness notifications cleanly — either
finding events to drain, or quietly continuing when the FD has no
pending events.
What did you see instead?
BlockingIOError raised from the Cython completion-queue poller and
propagated to asyncio's exception handler, logged at ERROR level
multiple times per second under load:
ERROR:asyncio:Exception in callback functools.partial(
<bound method PollerCompletionQueue._handle_events of
<grpc._cython.cygrpc.PollerCompletionQueue object at 0x...>>,
<uvloop.Loop running=True closed=False debug=False>)
handle: <Handle functools.partial(...)>
Traceback (most recent call last):
File "uvloop/cbhandles.pyx", line 61, in uvloop.loop.Handle._run
File "src/python/grpcio/grpc/_cython/_cygrpc/aio/completion_queue.pyx.pxi", line 148,
in grpc._cython.cygrpc.PollerCompletionQueue._handle_events
BlockingIOError: [Errno 35] Resource temporarily unavailable
Identical trace fires on the default _UnixSelectorEventLoop:
File "/opt/.../python3.11/asyncio/events.py", line 84, in _run
File "src/python/grpcio/grpc/_cython/_cygrpc/aio/completion_queue.pyx.pxi", line 148,
in grpc._cython.cygrpc.PollerCompletionQueue._handle_events
BlockingIOError: [Errno 35] Resource temporarily unavailable
So the bug is not loop-specific it's grpcio's own C-extension poller.
Anything else we should know about your project / environment?
What version of gRPC and what language are you using?
What operating system (Linux, Windows,...) and version?
errno 35(EAGAIN/EWOULDBLOCK)errno 11What runtime / compiler are you using (e.g. python version or version of gcc)
_UnixSelectorEventLoop)What did you do?
Used
grpc.aioas the inference-client transport in a productionbid-agent service. Topology per process:
/quoterequests. Each request fires 1-2 unary
PredictRPCs plus aPredictfor win-price prediction via agrpc.aiochannel pool._UnixSelectorEventLoopin aseparate thread) handles model loading, which also makes
LoadModelRPCs through the same channel-manager singleton.Channel options:
What did you expect to see?
grpc.aioto handle FD-readiness notifications cleanly — eitherfinding events to drain, or quietly continuing when the FD has no
pending events.
What did you see instead?
BlockingIOErrorraised from the Cython completion-queue poller andpropagated to asyncio's exception handler, logged at ERROR level
multiple times per second under load:
Identical trace fires on the default
_UnixSelectorEventLoop:So the bug is not loop-specific it's grpcio's own C-extension poller.
Anything else we should know about your project / environment?