stumbled upon it recently while optimizing parquet writes. It worked flawlessly and 10-20x'd my throughput