So does this mean the test was fair? If some libraries made assumptions or had default settings that by luck worked for your particular dataset does that mean they were quicker in converting to parquet?

Would it not have been better to use a dataset that didn't cause or be affected by these issues?