What was reported:
Issue milvus-io/pymilvus#2358 (closed, 8 comments): After upgrading grpcio to 1.68.0, multi-threaded concurrent DQL requests fail with "failed to connect to all addresses". Rolling back to grpcio 1.60.0 makes the errors disappear. Reproducible under concurrent search load, so it looked like a pymilvus bug at first.
What works:
Upstream grpcio bug, not pymilvus. Maintainer XuanYang-cn traced it to grpc/grpc#39894 and confirmed it is fixed in grpcio 1.71.x, 1.72.x and 1.73.x. The fix on the pymilvus side: upgrade grpcio to 1.71 or newer, but below 2.0, in your requirements and the concurrent connection failures go away. Newer pymilvus releases (from 2.6.0rc159 on) also loosened their grpcio pin so you stop getting forced onto the broken 1.68 line. If you are stuck on an older pymilvus for other reasons, grpcio 1.60.0 is the last known-good fallback.