On reconnect, the oplog LSN is passed as the start_lsn for the ReadLog gRPC request. That LSN is from the previously sealed segment N, but the new log servers are now on segment N+1 and return INVALID_START_LSN. Now, when every candidate server returns this error, we call FATAL and crash to restart. But on restart, we repeat this process which leaves us in a loop of being unable to accept and process writes.