sql-sentinel – Data & Science skill for AI agents

sql-sentinel

Overview

A static-analysis skill that audits SQL for the cost & performance anti-patterns that dominate warehouse bills — SELECT *, full-table scans, non-sargable predicates, Cartesian joins, the NOT IN NULL trap, and 15 more. It scores warehouse query health 0-100 (A-F) and outputs a prioritized cost-reduction plan, each finding with a why, a concrete fix, and an estimated savings.

Built for analytics engineers (dbt, Looker), data platform teams running FinOps / "reduce cloud spend" initiatives, and anyone reviewing a SQL pull request before it hits production. Works across BigQuery, Snowflake, Redshift, and Postgres. Zero dependencies, MIT licensed.

The executable engine and full rule set live in the source repository: https://github.com/takeaseatventure/sql-sentinel. Treat that repository as third-party executable code.

When to Use This Skill

• A user writes or reviews a query for BigQuery, Snowflake, Redshift, Postgres, or Spark SQL.
• A user asks "why is this query so slow?" or "why is my warehouse bill so high?"
• A user is about to promote a dashboard query or dbt model to production.
• A data engineer wants a second pair of eyes before a code review or a cost-optimization sweep.
• A team is running a "reduce cloud spend" or FinOps initiative.

How It Works

The engine splits a SQL script into statements (honoring quotes and comments), runs 20 rules over each statement, scores health 0-100 weighted by severity (critical 25, high 12, medium 5, low 1), and returns a prioritized cost-reduction plan.

Step 1: Run the audit

Install or clone the source repository only after choosing a reviewed commit, tag, or release to trust. Do not run code from a mutable default branch just because this skill links to it:

git clone https://github.com/takeaseatventure/sql-sentinel.git
cd sql-sentinel
git checkout <reviewed-commit-or-tag>
node scripts/sql-sentinel.js path/to/query.sql

Or programmatically:

const { auditSql } = require('./scripts/sql-sentinel');
const report = auditSql(yourSqlString, { dialect: 'bigquery' });
console.log(report.healthScore);      // 0-100
console.log(report.grade);            // 'A' | 'B' | 'C' | 'D' | 'E' | 'F'
console.log(report.prioritizedPlan);  // array, worst findings first

Step 2: Read the prioritized plan

The output leads with critical findings (Cartesian joins, mass DELETE) and descends to low-severity style issues. Each finding explains *why* it costs money and *how* to fix it.

Examples

Example 1: A messy dashboard query

SELECT DISTINCT *
FROM user_events, raw_logs
WHERE LOWER(event_name) LIKE '%signup%'
  AND user_id NOT IN (SELECT id FROM deleted_users)
ORDER BY created_at;

Illustrative candidate findings (no engine was run and no score/cost was measured):

• CRITICAL: comma-join without a join predicate produces a Cartesian product; actual impact needs plan/cardinality evidence
• HIGH: SELECT * may read unnecessary columns depending on engine/plan
• HIGH: leading-wildcard LIKE '%signup%' defeats indexes
• HIGH: LOWER(event_name) defeats indexes (non-sargable)
• HIGH: NOT IN (SELECT ...) — NULL semantics hazard
• MEDIUM: SELECT DISTINCT dedup cost
• MEDIUM: ORDER BY without LIMIT sorts the full result

Example 2: A clean, sargable query

-- Illustrative bounded query, not a measured score or proof of no findings
SELECT id, email, created_at
FROM users
WHERE created_at >= TIMESTAMP '2026-01-01'
  AND created_at <  TIMESTAMP '2026-02-01'
ORDER BY id
LIMIT 100;

Rules are heuristic candidates: expression/trigram indexes, partition pruning, small tables and intended CROSS JOIN/UNION semantics can change the verdict. Preserve result equivalence (including NULL/duplicate behavior), inspect EXPLAIN/warehouse dry-run and actual billing before recommending changes. EXPLAIN ANALYZE executes queries and needs separate data/cost authorization. A health score is not measured savings.

The 20 rules (ruleset v1.0.0)

| Rule | Severity | Catches |

|---|---|---|

| SQL001 | high | SELECT * full column scan |

| SQL002 | critical | No WHERE → full table scan |

| SQL003 | high | LIKE '%term' non-sargable |

| SQL004 | high | Function on column kills index |

| SQL005 | critical | CROSS JOIN / comma-join |

| SQL006 | medium | SELECT DISTINCT dedup cost |

| SQL007 | medium | ORDER BY without LIMIT |

| SQL008 | high | NOT IN (SELECT ...) NULL trap |

| SQL009 | medium | Implicit type cast |

| SQL010 | low | Many ORs (use IN/UNION) |

| SQL011 | medium | COUNT(DISTINCT) at scale (use HLL) |

| SQL012 | low | LIMIT without ORDER BY |

| SQL013 | medium | Scalar subquery in SELECT |

| SQL014 | medium | 5+ JOINs broadcast/spill risk |

| SQL015 | high | Fact table, no partition filter |

| SQL017 | low | String concat in SELECT |

| SQL018 | medium | Window OVER () no PARTITION |

| SQL020 | critical | DELETE/UPDATE without WHERE |

| SQL021 | low | SELECT * in EXISTS/IN |

| SQL022 | medium | UNION vs UNION ALL |

Run the test suite to verify each rule fires on real SQL:

cd scripts && node test.js   # 26 tests, zero dependencies

Limitations

• This is a static analyzer. It finds anti-patterns in the *text* of SQL; it does not read query plans, row counts, or billing. A flagged query on a 100-row table is cheap; the same query on a billion-row table is the problem the rule exists to prevent.
• The fact-table heuristic (SQL015) keys off table *names* (*_events, *_log) and is advisory, not definitive.
• The intended analyzer reads SQL text rather than executing it; review the pinned engine/dependencies before running and keep private SQL local unless external processing is authorized.