This talk revisits merge trees, a basic topological descriptor that records connectivity of sublevel sets of a scalar function. We introduce an interleaving distance between two merge trees and establish its stability to perturbations of the function. We show that this distance is never smaller than the bottleneck distance between 0-dimensional persistence diagrams of the function. On the computational side, we consider a distributed representation of merge trees that not only improves their parallel computation, but also supports parallel analysis. As an example, we show how to extract a prescribed levelset component of the function with minimum communication.