The kRollbackOplogSocketTimeout was raised to 10min from 30s in SERVER-26106, which may be too long to work with certain type of fault injections in tests. For example, BF-32624 injected a network partition error by dropping all packets on the TCP connection, where the only way to get out of it is timeout and retry. However, we don't usually wait 10min long in tests before we judge the node to be hung.

 
It's not ideal to change the server code to work around some testing limitation but I don't want to raise the timeout in the tests because that will make the passthroughs more time consuming.