Lang sql

lang-sql

This skill defines rules for writing correct, maintainable, and production-safe SQL — covering schema design, migrations, queries, and indexing.

Reference Guide

• PostgreSQL documentation: https://www.postgresql.org/docs/current/
• Use EXPLAIN ANALYZE to verify query plans before merging slow queries
• TIMESTAMPTZ (with timezone) over TIMESTAMP for any datetime column

Architecture Principles

Schema Design

• Use SERIAL PRIMARY KEY for surrogate keys
• Prefer TEXT over VARCHAR(n) unless a hard limit is meaningful (e.g. VARCHAR(45) for IP addresses, VARCHAR(500) for tokens)
• Use TIMESTAMPTZ NOT NULL DEFAULT NOW() for audit timestamps (created_at, computed_at)
• Nullable columns should be explicit: use DEFAULT NULL when adding optional columns
• Use UNIQUE constraints on the table, not just application-level validation
• Use CHECK constraints to enforce domain rules at the DB level (e.g. state machines, enum-like values)

State Machines in SQL

Model lifecycle states as a VARCHAR column with a CHECK constraint listing all valid values:

ADD COLUMN state VARCHAR(20) NOT NULL DEFAULT 'created'
    CHECK (state IN ('created', 'answering', 'submitted', 'completed', 'expired'));
COMMENT ON COLUMN quizzes.state IS 'Quiz lifecycle: created → answering → submitted → completed (or expired)';

Document the transition flow in a COMMENT ON COLUMN.

Foreign Keys

• Add ON DELETE CASCADE on child tables when child rows are meaningless without the parent
• Remove FK constraints intentionally when data preservation matters more than referential integrity (e.g. soft-delete QCMs while preserving quiz history) — always document this trade-off in the migration comment
• Never silently drop an FK; write a comment explaining why

JSON Columns

• Use JSONB (not JSON) for structured data that may be queried or indexed
• Only use JSONB when the structure is variable or too large for normalized columns (e.g. scores, question_sequence)
• Avoid JSONB for fields that are queried by value — normalize them instead

MUST DO

• One migration = one concern: each SQL file addresses a single schema change
• Number migrations sequentially: NN-description.sql (e.g. 21-add-ip-address.sql)
• Write a comment header explaining why the migration exists, especially for non-obvious changes
• Use IF EXISTS / IF NOT EXISTS on DROP and CREATE INDEX in migrations to make them idempotent
• Use partial indexes when a condition is almost always true (e.g. WHERE session_token IS NOT NULL)
• Use ON CONFLICT ... DO UPDATE (upsert) instead of separate SELECT + INSERT/UPDATE
• Use CTEs (WITH) to make complex queries readable; avoid deeply nested subqueries
• Use COALESCE to handle NULLs explicitly in calculations rather than relying on implicit NULL propagation
• Use EXTRACT(EPOCH FROM interval) * 1000 for millisecond durations from timestamp deltas
• Use LAG() window function for computing deltas between consecutive rows (e.g. answer timing)
• Parameterize all queries — never interpolate user input into SQL strings

MUST NOT DO

• Do not use SELECT * in application queries — always list columns explicitly
• Do not add columns without a DEFAULT on a large live table (it locks the table in old PG versions)
• Do not use TIMESTAMP without timezone — always use TIMESTAMPTZ
• Do not drop a foreign key without a comment explaining the trade-off
• Do not use JSON — always JSONB
• Do not put business logic (scoring, state transitions) in SQL — keep it in the application layer
• Do not use COUNT(*) to check existence — use EXISTS (SELECT 1 FROM ...) instead

Coding Style

Formatting

• Keywords in UPPERCASE: SELECT, FROM, WHERE, INSERT, ON DELETE CASCADE, etc.
• Table and column names in snake_case
• Indent continuation lines by one tab/4 spaces
• One column per line in CREATE TABLE and ALTER TABLE ADD COLUMN blocks

Migration Files

-- Migration: Short description of what this migration does
--
-- Purpose: Why this change is needed
-- Impact:  What tables/columns/indexes are affected
-- Trade-off: Any integrity or performance trade-off (if applicable)

ALTER TABLE quizzes
ADD COLUMN ip_address VARCHAR(45) DEFAULT NULL;

Indexes

-- Partial index: only index rows where token is set
CREATE INDEX idx_quizzes_session_token ON quizzes(session_token) WHERE session_token IS NOT NULL;

-- Composite index for analytics time-range queries
CREATE INDEX idx_answers_timing ON answers(quiz_id, answered_at);

Upsert Pattern

INSERT INTO answers (quiz_id, question_id, choice_id)
VALUES ($1, $2, $3)
ON CONFLICT (quiz_id, question_id)
DO UPDATE SET choice_id = EXCLUDED.choice_id;

Window Function for Timing

WITH ordered_answers AS (
    SELECT
        question_id,
        answered_at,
        LAG(answered_at) OVER (ORDER BY answered_at, id) AS prev_answered_at
    FROM answers
    WHERE quiz_id = $1
)
SELECT EXTRACT(EPOCH FROM (
    answered_at - COALESCE(prev_answered_at, (SELECT created_at FROM quizzes WHERE id = $1))
)) * 1000
FROM ordered_answers
WHERE question_id = $2;

Quality Standards

• Every migration must be tested on a local DB before committing
• Migrations are not reversible by default — write a separate rollback file only when rollback is planned
• All indexes on high-write tables must be justified (indexes slow writes)
• CHECK constraints must cover all valid values of an enum-like column
• IP address columns use VARCHAR(45) to support both IPv4 and IPv6

Agent Behavior

• When asked to add a column, always produce a numbered migration file in conf/initdb/
• When modifying a CHECK constraint, drop the old one first (DROP CONSTRAINT IF EXISTS) then add the new one
• When writing a JOIN query, always use explicit INNER JOIN / LEFT JOIN — never implicit comma joins
• When unsure whether a query will be slow, suggest EXPLAIN ANALYZE before shipping
• Never generate a migration that drops a column or table without first asking for confirmation