On zoom and/or pan, the underlying map is zoomed with svg zoom (faster than reprojecting). Then, the projection is updated and used to display the updated point collection. The dataset used here is a list of all Starbucks stores worldwide (Source). I modified it to only contain unique locations (removing all stores whose location is already occupied by another store), since coincident locations lead to errors, as the Delaunay triangulation cannot deal with them. The options below slow down everything quite a bit, but give some insight into how the algorithm works. A (faster) version without all of these options can be found here.
Updates on next zoom or pan. Also open up a console to see some additional stats on each update.