ADR-0024: Use typed input/output adapters around one transformation core¶

• Status: Accepted — owner-approved target; incomplete integration.
• Recorded: 2026-09-29. This is the recording date, not an invented original approval date.
• Authority and implementation boundaries: see Evidence below and the register conventions.

Context¶

CSV-only internals would lose native null/decimal/type provenance and duplicate transformation semantics when database/Parquet routes arrive.

Decision¶

Target four input families (CSV, Parquet, Trino query results, PostgreSQL query results) and three outputs (CSV, Parquet, PostgreSQL SQL script) through common transformation/validation. Bind source snapshots and explicit output schema/format before execution. SQL output is a file, not database writes.

Reuse scalar encoders and existing optional extras, but do not weaken synthetic-only writers to emit mixed-origin results. Do not use a lossy CSV roundtrip as the identity of a native source. Unsupported types and routes fail closed.

Alternatives and consequences¶

A separate engine per format duplicates safety and mapping semantics. Relabelling a Parquet-encoded internal capture as the original source loses adapter/query identity. Private CSV-named functions and injected fictional drivers are transitional implementation details, not the final public API.

Twelve-route acceptance is a target, not a claim of completion. Named cursors, worker isolation and decoded-byte limits alone do not prove hard wire/RSS containment or scale.

Evidence¶

Sources: resumption architecture, policy input/output contract, implementation map. Executable anchors: Parquet, SQL output, query capture.

Revisit when¶

Choose public adapter contracts only with typed, end-to-end evidence; do not treat private function names as permanently accepted architecture.