Just be careful with figures from papers they tend to be rather asymptotic. IBLT like schemes usually start with ~32-bits per difference in overhead for a checksum which is pretty bad when otherwise 30-bit members are fine >100% overhead before you even get to the overheads needed to achieve correct reconstruction. 
Some papers simply leave this overhead out of their figures, though I don’t recall if the rateless paper did that.
(FWIW, you can use minisketch in a kind of quasi rateless way by just dynamically sending more until the other side could recover– almost all the computation from a partial one can be conserved, as you don’t get to the expensive and non-reusable root finding step until you’re almost certain to have a correct decode, so long as you’re willing to take one or two extra elements overhead)