Skip to content

ScanIterator: misaligned per-column chunk grids rejected — only 1-vs-N chunking supported #221

Description

@dfa1

Found by Raincloud triage round 3 (#205) on two datasets the Rust oracle reads fine. ScanIterator.buildChunks enforces "mixed per-column chunking beyond 1-vs-N is not supported":

VortexException: scan: column 'PM2.5' has 4 flats but the widest column has 10; ...
    at ScanIterator.buildChunks(ScanIterator.java:144)

Two severity tiers, same guard:

  1. Nested boundariesemotions-dataset-for-nlp (416,961 rows): label has 4 chunks [131072 ×3, 23593], text has 26 chunks (25×16384 + 7209); 8×16384 = 131072, so every coarse boundary is also a fine boundary. A boundary-nesting generalization of the 1-vs-N rule (slice the coarse column at the fine grid) fixes this tier.
  2. Disjoint gridsuci-beijing-multi-site-air-quality (420,768 rows): numeric columns chunk [131072 ×3, 27552] while station chunks [40960, 49152, 49152, 49152, 49152, 40960, 49152, 49152, 40960, 2976]; 40960+49152+49152 = 139264 ≠ 131072 — boundaries do not nest. This tier needs true chunk realignment/splicing during scan (emit chunks at the merged boundary grid, slicing every column to each window).

Worth checking how the Rust scan planner merges per-column chunk grids before designing the fix. Matrix entries: both datasets pinned as gaps referencing this issue.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions