본문으로 건너뛰기

ADR 0016 - Ordering and stability

  • Status: accepted
  • Date: 2026-05-08
  • Spec target: XTL 0.1 draft
  • Affects: language.md, evaluation.md

Context​

Three ordering questions have lurked under the spec without a normative answer. Each one is a portability hazard if two conformant implementations make different choices:

  1. @sort stability. evaluation.md defines the directive pipeline (filter → sort → top) but never says whether @sort preserves the relative order of rows that have equal sort keys. Two implementations could sort identically by key and still differ on tie order.
  2. File group output order. When output_file_pattern produces multiple files (one per Customer, etc.), the spec doesn't say in what order those files appear in the result list / zip. The reference impl uses first-seen order today; that is observable behavior nobody pinned.
  3. Sheet group order within a file. When a sheet template has a group key (Region etc.) and produces multiple sheets per file, the order is currently order.sort() — lexicographic on the canonical-string-form group key. That is inconsistent with the file-group rule and depends on the JS default sort, which produces a different relative ordering than file groups for the same data.

Pre-1.0 is the right window to lock these.

Considered Options​

@sort stability​

A. Stable (chosen). Equal sort keys keep their input order. JS Array.sort is stable since ES2019. Easy to implement; matches author intuition.

B. Unstable, implementation-defined. Cost: silently breaks data flows that rely on a stable secondary order (e.g. sort by region, then expect customers to appear in source order within each region).

C. Lexicographic tie-breaker on first column. Cost: surprises authors; requires them to know the rule.

File group order​

D. First-seen (chosen). When the engine walks source rows in order, the first row whose group key produces filename X writes that group "first." Reproducible from the source data without further sort.

E. Sorted by canonical-string-form filename. Cost: every run sorts files lexicographically — predictable but doesn't match the source's own order; authors expect "first row's group goes first."

Sheet group order within a file​

F. First-seen, matching file groups (chosen). Reuses the same rule; once authors learn it for files they know it for sheets.

G. Lexicographic by joined key (status quo). Inconsistent with file groups, locale-undefined.

Decision​

@sort is stable​

Implementations MUST preserve the relative order of rows whose @sort key compares equal. With multiple @sort directives, the first directive is the primary sort key; subsequent directives are tiebreakers in the order they appear. Source order is the final tiebreaker. This matches the convention of Excel "Sort by … then by …" and SQL ORDER BY a, b (the first column is primary).

File and sheet groups use first-seen order​

When output_file_pattern evaluates to multiple distinct filenames, the file groups are emitted in the order in which the engine first encounters each filename while walking source rows.

When a sheet template's group keys produce multiple sheets within a file, those sheets are emitted in first-seen order over the file group's row list.

The single-source iteration order is the source's natural row order (source_table reading top-to-bottom). With multi-source data (ADR-0012) this rule applies to the primary source's rows.

Spec text additions​

  • language.md "Sort" gains: "@sort is stable. Equal sort keys preserve source order. Multiple @sort directives apply in order with later sorts stable relative to earlier ones."
  • evaluation.md "Render Phases" gains a new "Ordering" subsection:
    • File groups: first-seen.
    • Sheet groups within a file: first-seen.
  • The implementation's sheet-group lexicographic sort (order.sort() in grouper.ts:groupByKeysOrdered) is replaced with insertion order.

Consequences​

  • Single-source templates that sort by a column are unaffected (JS default Array.sort is already stable).
  • Templates that relied on the implementation's lexicographic sheet- group ordering will see a behavior change: sheets now appear in the order their first matching source row appeared. This is a deliberate, signalled break for the 0.x window.
  • Multi-source templates (ADR-0012) inherit the same rule: the primary source's iteration order drives output order; named sources contribute to aggregates and joins but do not affect file / sheet output order.
  • The spec gains predictability for ZIP packaging, audit, and diffing across runs.
  • This ADR does not introduce author control over output order. Authors who need a specific order can sort their source data beforehand or use a sortable group-key column.

References​

  • ADR-0007: empty value definition (empty group keys still produce one group per distinct empty representation; ordering by first- seen).
  • ADR-0009: comparison and string coercion (canonical-string-form drives group-key identity; ordering does not depend on Unicode collation).
  • ADR-0012: multi-source data model (primary source drives ordering).
  • spec/evaluation.md "Render Phases", "Group Keys".
  • spec/language.md "Sort".