Mixing RLE_DICTIONARY and other column encodings in pyarrow parquet

[bkief]

Describe the bug, including details regarding any error messages, version, and platform.

The ValueError at

arrow/python/pyarrow/_parquet.pyx

Lines 1360 to 1367 in aaeff72

	'DELTA_BYTE_ARRAY': ParquetEncoding_DELTA_BYTE_ARRAY,
	'RLE_DICTIONARY': 'dict',
	'PLAIN_DICTIONARY': 'dict',
	}.get(encoding_name, None)
	if enc is None:
	raise ValueError(f"Unsupported column encoding: {encoding_name!r}")
	elif enc == 'dict':
	raise ValueError(f"{encoding_name!r} is already used by default.")

will raise anytime any column is custom encoded with a dictionary method. This makes it impossible to mix a dictionary encoded column with something like DELTA_BINARY_PACKED. I understand this is to prevent duplication of use_dictionary. Would it be okay to move this ValueError to the calling function instead? Does anything at the C++ level prevent this?

arrow/python/pyarrow/_parquet.pyx

Lines 1971 to 1972 in aaeff72

	elif isinstance(column_encoding, str):
	props.encoding(encoding_enum_from_name(column_encoding))

Component(s)

Parquet, Python

[mapleFU]

I didn't fully catch what you mean, do you mean:

1. By default, using dictionary encoding
2. Using Delta as fallback encoding. If dictionary is too large, a row-group will fallback to use Delta

Am I right?