Skip to content

CLAUDE.md

Latest commit

 

History

History
24 lines (18 loc) · 1.64 KB

CLAUDE.md

File metadata and controls

24 lines (18 loc) · 1.64 KB
Raw

all instructions see AGENTS.md

before doing anything at the untangle function, read the two papers for Neumann Unnesting

Neumann Query Unnesting

Papers (MUST READ at start of every session)

  • /home/carli/projekte/memcp/papers/Unnesting-Arbitrary-Queries.pdf (Neumann/Kemper NK15, BTW 2015)
  • /home/carli/projekte/memcp/papers/neumann-improving-unnesting-btw2025.pdf (Neumann BTW 2025)

Core Principles

  1. NO fallback plans for inline evaluation. Every query is unnestable per Neumann. If unnest_subselect returns nil, it's a BUG — fix the unnesting, don't add inline workarounds.
  2. Every inner_select is unnestable. The old build_scalar_subselect, build_in_subselect, build_exists_subselect were removed intentionally. They must NOT be brought back.
  3. Clean architecture only. No workarounds, no magic strings, no special-case hacks. The flat table list + scoped group stages + partition stages IR is the correct abstraction.
  4. Tests green through bugfixing, not code destruction. Identify WHY a case fails, trace the IR, find the root cause, fix it cleanly. Do NOT disable tests, mark them noncritical, or add fallback paths.

Current Regression Status (60 regressions vs master)

The removed inline subselect functions handled cases that unnest_subselect currently returns nil for:

  • us_outer_in_fields → nil (outer refs in SELECT list of subquery)
  • us_single_tbl = false → nil (multi-table subselects)
  • UNION ALL subqueries → nil
  • Various computed expressions in subselect fields

These are NOT unsolvable — Neumann covers all of them. They need proper implementation of the push-down rules, not fallback plans.