I’ve tried reading streamed parquet via PyArrow with Duck, and it’s been pretty promising. Depending on the query, you won’t need to download everything off HTTP.