Skip to content

A string column whose first 8192 rows are constant gets a dict layout Rust does not write #722

Open

Description

@dfa1

Rust's dict layout runs its probe compressor over the first 8192 rows and takes the dict layout only when the result is a Dict (vortex-layout/src/layouts/dict/writer.rs:169-182, 0.91). A constant sample compresses to Constant, so Rust falls back to plain chunked flats.

For Utf8/Binary, GlobalDictWriter.detect uses DictColumnState.isVarBinDictCandidate ("not more than half distinct") instead, which admits a sample of 0 or 1 distinct values.

Found on Raincloud coig: the first 8192 rows of instruction are all "", so Java writes a global dict layout and the column is 8.8% bigger than Rust's (5.27 MB vs 4.85 MB from vortex-jni 0.91). With withGlobalDict(false) Java writes 4.84 MB in 13 chunks, Rust's size and chunking. The constant split columns of coig and mnist get a dict layout for the same reason.

Fix: a sample with at most one distinct non-null value is not a candidate. The full parity fix is to ask the cascade probe for strings too, as primitives do, once the string scheme picks Dict where Rust's does (it picks OnPair today).

Test first: a vortex-jni cross-check on a string column whose first 8192 rows are constant and the rest varied, asserting both writers pick the same layout.

Activity

  1. added
    bugSomething isn't working
    on Nov 10, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions