---

name: sql-db-architecture description: Understand the toy relational database architecture, crate organization, data flow, and implementation patterns. Use when exploring the database codebase, understanding query execution, debugging storage/indexing, or planning new features. Keywords: sql-database, database, crate, storage, executor, planner, parser, catalog, WAL, buffer pool, index, architecture

sql-database Architecture Expert

Provides deep understanding of the toy relational database implementation, crate boundaries, and system design.

When to Activate

Use this skill when:

• Exploring how SQL queries flow through the system
• Understanding storage layer (heap files, slotted pages, WAL)
• Debugging parser → planner → executor pipeline
• Working with indexes (B+Tree, Hash, Bitmap, Trie)
• Navigating crate dependencies and boundaries
• Planning new features that span multiple crates

System Overview

sql-database is a minimal RDBMS in Rust implementing:

• SQL subset: CREATE/DROP TABLE/INDEX, INSERT, SELECT, UPDATE, DELETE
• Storage: slotted pages + buffer pool + WAL (redo-only)
• Execution: Volcano model with SeqScan/IndexScan operators
• Indexes: B+Tree (sled), Hash, Bitmap (roaring), Trie

Instructions

1. Identify the Query

When user asks about functionality, map to crates:

"How does SELECT work?" → parser → planner → executor → storage/index → buffer

"Where are tables created?" → catalog (metadata) → storage (heap allocation) → WAL (logging)

"How are indexes used?" → planner (index selection) → executor (IndexScan) → index crate

2. Crate Responsibility Map

Use this mapping to locate relevant code:

3. Data Flow Patterns

Query Execution (SELECT):

CopySQL string
  → parser (sqlparser-rs → AST)
  → planner (logical plan → physical plan with index selection)
  → executor (Volcano operators: Scan → Filter → Project)
    → storage/index (fetch rows via RecordId)
      → buffer (page cache lookup/fetch)
  → REPL (format RecordBatch with tabled)

Write Path (INSERT/UPDATE/DELETE):

CopySQL string
  → parser → planner → executor
  → WAL (append WalRecord, fsync)
  → storage (heap table modification)
  → index (maintain all indexes on affected columns)
  → buffer (mark pages dirty)

Recovery:

CopyStartup
  → WAL replay (read WalRecord log)
  → storage (re-apply operations)
  → catalog (restore table/index metadata)

4. Code Navigation Strategy

Finding implementations:

• Trait definitions: Search crate root (e.g., executor/lib.rs for Exec)
• Concrete operators: Look in <crate>/src/<operator>.rs
• Tests: Adjacent .rs files or tests/ directory
• Integration: testsupport/ for end-to-end scripts

Dependency order (from CLAUDE.md):

Copycommon ← types ← expr
         ↓
       parser
         ↓
      catalog
         ↓
  buffer ← storage ← wal
         ↓
      planner
         ↓
       index
         ↓
     executor
         ↓
       repl

5. Key Design Principles

From the design doc:

1. Tiny interfaces - Prefer small traits (Exec, Index, HeapTable, Pager)
2. No implicit coercion - Same-type comparisons only (v1)
3. WAL-first writes - (1) append WAL (2) fsync (3) apply to storage
4. Rule-based planning - Predicate pushdown + index selection
5. Single writer - No transactions/MVCC in v1
6. Snapshot tests - Use insta for query output verification

6. Response Format

When explaining architecture:

Component Overview:

• Purpose in 1 sentence
• Key types/traits
• Dependencies (what it calls)
• Dependents (what calls it)

Code References:

• Use crate/path/file.rs:line format
• Link related components
• Show trait → impl relationships

Examples:

• Provide concrete SQL queries
• Show expected output format
• Reference existing tests when available

Project Context

• Language: Rust (workspace with 13 crates)
• Build: cargo check/test/fmt/clippy
• Coverage: scripts/coverage.sh (llvm-cov)
• Testing: Unit tests inline, integration in tests/, snapshots with insta
• Dependencies: Pinned in workspace Cargo.toml with workspace = true

Common Questions

"How does index selection work?" → Check planner/ for rules matching WHERE predicates to available indexes

"Where is Row serialized?" → storage/ uses bincode for tuple layout in slotted pages

"How does the buffer pool work?" → buffer/ implements LRU cache over PageId, backed by file segments

"What SQL is supported?" → See SKILL.md header or parser/ for DDL/DML subset

"How to add a new operator?" → Implement Exec trait in executor/, wire into planner physical node generation

Tool Usage

• Grep: Find trait implementations, error types, specific SQL keywords
• Read: Examine crate root lib.rs, trait definitions, test fixtures
• Glob: Locate all operators (executor/**/*.rs), test files (**/tests/*.rs)

---

For detailed API signatures and implementation notes, see REFERENCE.md.