CSV, JSONL, Parquet: choose the file by the job

Dataset format is not a taste question. It is a promise about how the next person or job can read the data.

CSV is fine for small, flat tables that humans may inspect. JSONL is useful when records arrive one at a time or carry nested payloads. Parquet is usually the better fit for feature tables that will be scanned repeatedly by analytics or training jobs.

The builder move is to name the downstream job first, then pick the simplest format that supports it.