Critical Description
Currently when getMeta receives an error that is not network timeout type, it repairs network connection and continues the loop that calls readFromClient. This is not correct. After the network connection is repaired, there is nothing to read from the connection, and readFromClient will always time out in 2 minutes. Within the 2 minutes nothing else can get done and replication will be seemingly stuck.
The right thing to do is to let getMeta abort with error right away after it repairs connection.