[viethqb]

Summary

• Plan: adds docs/superpowers/plans/2026-10-24-p0-m1a-catalogs.md, the implementation plan for P0/M1a: Iceberg catalogs and single-node reads #16 (M1a: Iceberg catalogs and single-node reads). It has 7 tasks:
  1. Docker test environment: PostgreSQL, RustFS, Iceberg REST catalog, and Spark 4 + Iceberg 1.10 generating the fixtures
  2. egret-engine-meta (PostgreSQL metadata store)
  3. egret-engine-sql (CREATE/DROP/SHOW CATALOG parser)
  4. egret-engine-iceberg (DataFusion catalog, table and scan over iceberg-rust)
  5. egret-engine-catalog (registry, secret references, REST options)
  6. egret-engine-coordinator (engine, sessions, SQLSTATE), with pgwire reduced to the protocol
  7. Server wiring, end-to-end tests against Spark's results, and docs
• Roadmap: splits M1 into M1a (P0/M1a: Iceberg catalogs and single-node reads #16), M1b (P0/M1b: Engine scan planning and metadata caches #17) and M1c (P0/M1c: Snapshot diff API and credential refresh #18), all sub-issues of P0/M1: Iceberg single-node #8, and records the pre-work check M1 required.
  • The iceberg-rust evaluators and delete_file_index are pub(crate) at rev 5b82ab99 and on upstream main, so M1b needs a thin fork.
  • The new ObjectCacheProvide trait exists but is not wired into Table. M1b will consider wiring it in instead.

Verification

All code in the plan was built and run before it was written into the plan.

• Task replay: the tasks were replayed one by one on a clean worktree from main. After each task, fmt, clippy (-D warnings) and all tests passed. The final state has 71 tests: 61 run without Docker and 10 need the environment.
• Spark comparison: every fixture query (7 in total) returns the same rows over pgwire as in Spark. This covers positional deletes, deletion vectors (PUFFIN), and schema and partition evolution.
• Manual check: I ran the README example against the binary with psql. SELECT count(*) FROM lake.fixtures.deletion_vectors returns 772.

Known gap, to be added to #12: the engine prints timestamps with six fractional digits, while PostgreSQL omits a zero fraction. The fixture rendering follows the engine.

Refs #8, #16

🤖 Generated with Claude Code