I think it may make sense to continue developing and maintaining this code in the same place as other file format <-> Arrow serialization code and dataset handling routines (e.g. schema normalization). Under this scheme, libparquet becomes a link time dependency of libarrow_dataset