A write that fails may result in a call to setLastOpToSystemLastOpTime. We invariant that the last optime on a client never goes backwards, but the system optime can indeed go backwards after a rollback. If that happens we should just refuse to set the last optime on the client backwards, but we shouldn't crash.