How "huge" are these datasets? Did you build your own tooling to accomplish this?