Output Parquet File

How it fits

Typical backends

Caveats

• I/O contractEvery Parquet column is stored as a string, mirroring the incoming dataset; downstream readers must cast to typed values themselves.
• I/O contractA populated header must have non-empty, unique column names or the write aborts; when the header is empty, columns are named by their zero-based index derived from the widest row.
• I/O contractEach cell is taken positionally — rows shorter than the column count are padded with empty strings, and a dataset with zero rows or zero columns writes no file at all.
• Hard constraintcompression is matched case-insensitively against the supported codecs; an unsupported name aborts the write rather than falling back.
• State lifetimename_template substitutes a process-local write index via {index:06d} or {index}; the counter resets on restart and, without a placeholder, each write reuses one name and overwrites the prior file.