Description
Currently, the server is responsible for timing out RPC keepalive connections. This means that there can be a race where a client sends an RPC exactly at the same time that the server is about to close the connection due to keepalive. In this case, the client gets a "Connection reset by peer" trying to read its response, and the RPC fails.
I've seen this cause a failure of linked_list-test once.
Attachments
Issue Links
- is related to
-
IMPALA-6159 DataStreamSender should transparently handle some connection reset by peer
- Resolved
-
IMPALA-5557 Disable rpc_default_keepalive_time_ms
- Resolved