Motivation

This is a continuation of #1938. Other errors are not fatal to the server as a whole.

I observed this today in production (we are already running #1938). While connections breaking are much less rare than "unexpected eof", they still happened once:

Sep 19 14:56:13 api01 watchdog.sh[19879]: {"timestamp":"2024-09-19T14:56:13.360774Z","level":"DEBUG","message":"accept loop error","error":"Connection reset by peer (os error 104)","target":"tonic::transport::server::incoming"}

Solution

The solution is to add "connection reset" and "broken pipe" errors to the list of non-fatal errors that make the accept loop continue.