The current implementation assumes that there always exists a huge input that we want and can use to queue transactions on top of each other. We need to teach the faucet to consume other inputs to minimize the dust.