Apache Parquet's Dremel encoding treats all lists as variable-length, causing roughly 3× overhead when reading fixed-length lists like vector embeddings or 3D coordinates. The Hardwood Parquet library implements a fast path that inspects encoded definition and repetition level streams to detect when all lists in a data page share the same length. When detected, it bypasses Dremel record reconstruction, copies values page-wise, skips level array materialization, and computes offsets trivially. The detection uses SWAR bit-pattern matching for small list sizes (n≤8), stride-based bulk comparison for large sizes (n≥16), and a scalar fallback for the awkward 9–15 range. Results show 2.5× speedup for the column reader and up to 3.9× for the row reader on 768-element embeddings, approaching flat-column performance. Detection overhead is negligible (<2%) even in worst-case scenarios. The optimization ships in Hardwood 1.1.0.Beta1 as an opt-in flag.
