This is only intended as an interim solution and a last resort when we are faced with downtime and disrupted cluster operations. When we lose the majority of our servers and we still have enough servers to start a replica set (maybe including some quickly spawned arbiters), we can force a reconfiguration with only the surviving members.
First, we get the replica set configuration document:
> cfg = rs.conf()
Using printjson(cfg), we identify the members that are still operational. Let's say that these are 1, 2, and 3:
> cfg.members = [cfg.members[1] , cfg.members[2] , cfg.members[3]]
> rs.reconfig(cfg, {force : true})
By using force : true, we are forcing this reconfiguration to happen. Of course, we need to have at least three surviving members in our replica set for this to work.
It's important to remove the failing servers as soon as possible, by killing the processes and/or taking them out of the network, avoid unintended consequences; these servers may believe that they are still a part of a cluster that doesn't acknowledge them anymore.