captain kiyaku already does that. even i already did it in my experiment. it wouldn't make sense to generate faces that never can be seen.
yes, using a big array is no good idea. better store the voxels directly with the mesh chunks and write some dynamic chunk manager. then you also can more easily create new chunks if the player moves towards the current boundaries of the world.
but with a world as big as the surface of the earth i wonder about floating point precision issues. probably the player somehow would have to be moved to the coordinate origin again after a while to avoid problems. without him noticing.
and with realtime shadows...
