ADR-0001: Target Postgres only

• Status: Working
• Confidence: High
• Decided: 2026-07-27
• Last reviewed: 2026-07-27

Context

The features sqlb is built on are not evenly available across SQL dialects: LISTEN/NOTIFY for the change feed, jsonb aggregation for relation expansion, RETURNING on every mutation, ON CONFLICT, SKIP LOCKED, partial and GIN indexes, ILIKE. Supporting a second dialect means either dropping to the intersection or carrying per-dialect branches through the compiler, the schema DSL and the migration generator.

Decision

Target Postgres, and only Postgres. Keep a Dialect interface for placeholder style and identifier quoting, but do not pretend to be portable.

Consequences

Buys. Every feature can assume the best available primitive. Mutations return their rows in one round trip. The compiler stays small — one set of rules.

Costs. Unusable for MySQL or SQLite projects. Tests wanting an in-memory database need a real Postgres or the fake driver. LISTEN needs a session, so a pooler in transaction mode takes part of the change feed back (ADR-0019).

What would change our mind

• A concrete project we want to support is on MySQL and cannot move.
• The Dialect interface leaks — if compiling needs dialect knowledge outside placeholder rendering and quoting, the seam is in the wrong place.
• SQLite-in-tests becomes painful enough that the fake driver stops being enough.

Cost of change

Asymmetric. Narrowing is free; widening is close to a restart — every compiler assumption audited, replacements found for LISTEN/NOTIFY and jsonb expansion, every generated migration re-validated.

Revisions

• 2026-07-27 — Written.
• 2026-07-27 — Qualified the LISTEN/NOTIFY benefit: it does not hold through a transaction-pooling connection pooler.
• 2026-07-30 — Condensed.