You may want to give octosql a try. I was able to read 100GiB parquet files and aggregate them using less than 100MiB RAM and it is pretty fast too