Trying to figure out how to write partitioned parquet files using either pandas or using Pyarrow directly (which is used in the background by pandas). The goal is to have a partitioned file for our whole feature sets, such that we can access the features by their partition colums, for example - for particular `gm_id` and `product_id` pairs. This will hopefully reduce the amount of I/O that needs to be done by the workers during hyper parameter tuning.