Skip to content

About

These pages argue for the design of datarecord: what it chose, what it rejected and what each choice costs. The rules themselves are in the reference.

  • One record: why one interface reads a directory and a tree of layers, and what datarecord leaves to others.
  • The data model: why a dim is flat, a multi-dim is one struct column, a category is a dim and a result is an attribute.
  • Layers: why a layer is written once, restates an axis whole and overrides one value per label only along partial.
  • Storage: why parquet and DuckDB, why writing a manifest creates a revision, and why writers never share a file.
  • Open questions: what the design leaves unsettled, and the questions it has settled.