Le 02/07/2026 à 17:18, Russell Spitzer a écrit :
I find Phillpp's arguments here pretty strong, but I'd like to know more about how others think about the actual Parquet implementation. I'm not an expert on how the read code or definition levels work right now but it does seem to me that most downstream consumers of fixed-length numeric vectors are going to always see them as whole objects?
That question seems too vague to give a general answer, but in Arrow those would be represented as FixedList<T, N>, potentially annotated with a Fixed shape tensor extension type: https://arrow.apache.org/docs/format/CanonicalExtensions.html#fixed-shape-tensor
Regards Antoine.
