To get "sticky read your writes" behavior out of the internal replica set client, used by mongos and the shell, the internal client's code for replica set monitoring and server selection should accept as "primary" nodes reporting a larger electionId value than the most recently selected primary node.

Note that in 2.6 and 3.0, large clock skews (> 30 seconds) between the replica set member nodes may lead to clients that observe an election being unable to issue writes or primary reads. The solution would be to restart those clients.