It looks like both Dplython and pandas-ply are missing one of (what I think is) the core value propositions of dplyr: the ability to use the same abstractions on local data and on remote data, with execution against the remote source happening lazily such that the entire table doesn't need to be downloaded in order to run a filter locally.

(Of course, I may be biased in that I work on a commercial product which also has this characteristic.)