The mechanism in mongos to select a replica with secondary read preference does not do what the documentation says it does.

At http://docs.mongodb.org/manual/core/read-preference-mechanics/#member-selection it says that the algorithm takes the nearest replica (lowest ping time), and then computes a set of other replicas that are no further than the lowest ping time plus the value of --localThreshold. It then chooses one of these replicas at random.

Looking at the code in dbclient_rs.cpp, it takes --localThreshold as an absolute value and returns the first node that has a ping time less than that. If there is no such node, it returns a random node that matches the tags.

For comparison, I checked the Java driver and found that it implements what the documentation says.