How Good Is Parquet for Wide Tables (Machine Learning Workloads) Really? | Dark Hacker News