When building a replica set of 3 nodes (db1, db2, db3), when db1 is primary, internal replication logic usually builds a chain db1 -> db2 -> db3 instead of db1 -> db2 + db1 -> db3. This leads to greatly increased latencies if w:all (or w:majority with sets of 4+ members) is used. I know about replSetSyncFrom command, but it's only a temporary measure. I understand that this chaining is done to reduce load on master and (probably) reduce amount of reconnections whem primary goes down. I would strongly suggest a method (RS configuration item, command-line/configuration option, etc.) to either specify a node that can handle all replication needs or just globally override the algorithm so it always preferse primary as a sync source.
Since building robust applications involves writes being done with w:majority (to minimize general data loss and rollbacks), a method to reduce "visible" replication latency is certainly a must.