Drizzle

Drizzle ORM

CRITICAL: Your training data for Drizzle ORM is unreliable. APIs change between versions and memorized patterns may be wrong or deprecated. Before writing any code, you MUST use WebFetch to read the live docs:

WebFetch("https://orm.drizzle.team/llms.txt")

Do not proceed without fetching this URL first. Never assume an API or config option exists — verify against current docs.

Drizzle ORM is a lightweight, TypeScript-first ORM with a SQL-like query builder, schema-as-code migrations, and zero dependencies.

Key Capabilities

Built-in schema validation adapters — Drizzle includes first-party integrations with Zod, Valibot, TypeBox, ArkType, and Effect Schema via drizzle-zod, drizzle-valibot, etc. There is no need to manually write validation schemas that mirror your table definitions; these packages infer insert/select types directly from your Drizzle schema.

Built-in seeding — drizzle-seed provides deterministic, reproducible fake data generation with seedable PRNG, weighted distributions, and cross-dialect support. It replaces ad-hoc seed scripts or external packages like @faker-js/faker for database seeding workflows.

Best Practices

Call .$dynamic() to compose queries across functions. By default, Drizzle query builders restrict each clause method (where, limit, orderBy) to a single call — this mirrors SQL but breaks composable helper functions. You must call .$dynamic() to unlock multi-call mode. Use the dialect-specific generic types (PgSelect, MySqlSelect, SQLiteSelect) as parameter types for functions that accept a query builder to extend.

Relational queries (db.query.*) generate exactly one SQL statement — nested offset is not supported. Unlike Prisma's batched-query approach, Drizzle uses a single lateral-join query per db.query call. A consequence is that offset only works at the top level of a relational query; nested relation pagination must use where with cursor-based logic instead.

Use drizzle-kit push only for development, never production. drizzle-kit push applies schema changes directly without generating migration files. Teams that use push in production lose the auditable migration history needed for drizzle-kit migrate in CI/CD. Use drizzle-kit generate + drizzle-kit migrate for production deployments.

Never instantiate drizzle() per request in serverless environments. Each call can open a new connection pool. Export a module-level singleton and re-use it across invocations, following the same pattern as Prisma's PrismaClient singleton pattern.

drizzle-kit pull and push do not preserve hand-written SQL in migration files. If you manually edit a generated migration file (e.g., to add a partial index or a custom trigger), subsequent pull operations will overwrite those edits. Track custom migrations in separate numbered files and apply them alongside generated ones.

db.query.* relational queries require explicit relations() definitions. Unlike Prisma, Drizzle does not infer relations from foreign keys automatically. If you call db.query.users.findMany({ with: { posts: true } }) without a relations() export in the schema file, the query throws at runtime with a non-obvious error. Define relations explicitly for every table you want to query relationally.