---
name: api-database-edgedb
description: Graph-relational database with EdgeQL query language, code-first schema, link-based relations, computed properties, and fully typed TypeScript query builder
---

# Gel (formerly EdgeDB) Patterns

> **Quick Guide:** Gel (formerly EdgeDB) is a graph-relational database built on PostgreSQL. Define schemas in `.gel` files using SDL with types, links, and computed properties. Use `gel migration create` + `gel migrate` for schema changes. Query with EdgeQL (set-based, deeply nested shapes) or the TypeScript query builder (`e.select`, `e.insert`). Everything in EdgeQL is a set -- empty sets need explicit casts, and operations on sets produce Cartesian products. Use `global` variables with access policies for row-level security. The query builder requires a running database for code generation (`npx @gel/generate edgeql-js`).
>
> **Naming:** EdgeDB was rebranded to **Gel** in February 2025. The `edgedb` npm package, CLI, and `.esdl` extension still work via compatibility shims, but new projects should use `gel`, `@gel/generate`, and `.gel` files.

---

<critical_requirements>

## CRITICAL: Before Using This Skill

> **All code must follow project conventions in CLAUDE.md** (kebab-case, named exports, import ordering, `import type`, named constants)

**(You MUST run `npx @gel/generate edgeql-js` after every `gel migrate` -- the generated query builder is based on the database schema and becomes stale after migrations)**

**(You MUST cast empty sets explicitly (`<str>{}`, `<int64>{}`) -- bare `{}` is a syntax error because EdgeQL is strongly typed and cannot infer the type of an empty set)**

**(You MUST understand that all EdgeQL values are sets -- operations on multi-valued expressions produce Cartesian products, not element-wise results)**

**(You MUST pass the transaction object `tx` (not `client`) to ALL query `.run()` calls inside `client.transaction()` -- using `client` inside a transaction runs queries outside the transaction)**

**(You MUST NOT use volatile functions like `datetime_current()` in schema-defined computed properties -- use `datetime_of_transaction()` or `datetime_of_statement()` instead)**

</critical_requirements>

---

**Auto-detection:** Gel, gel, EdgeDB, edgedb, EdgeQL, edgeql, .gel, .esdl, dbschema, edgeql-js, createClient, e.select, e.insert, e.update, e.delete, e.params, gel migrate, gel migration, edgedb migrate, SDL schema, backlink, access policy, gel.toml, edgedb.toml

**When to use:**

- Defining graph-relational schemas with types, links, and computed properties
- Writing type-safe queries with EdgeQL or the TypeScript query builder
- Managing schema migrations with the built-in migration system
- Modeling complex relationships (multi links, backlinks, polymorphism)
- Implementing row-level security with access policies and globals

**Key patterns covered:**

- Client setup and connection (`createClient`, DSN, environment variables)
- Schema definition in SDL (types, properties, links, constraints, computed)
- EdgeQL query language (SELECT shapes, INSERT, UPDATE, DELETE)
- TypeScript query builder (`e.select`, `e.insert`, `e.update`, `e.delete`)
- Migrations workflow (`gel migration create`, `gel migrate`)

**When NOT to use:**

- Simple key-value storage (use a dedicated key-value store)
- Projects that need raw SQL as the primary interface (Gel uses EdgeQL; Gel 6+ adds native SQL support but EdgeQL is the primary interface)
- Environments where you cannot run the Gel server (it is not an embedded database)

**Detailed Resources:**

- For decision frameworks and quick reference, see [reference.md](reference.md)

**Core Patterns:**

- [examples/core.md](examples/core.md) - Client setup, schema definition, EdgeQL basics, migration workflow

**Query Builder:**

- [examples/query-builder.md](examples/query-builder.md) - TypeScript query builder (e.select, e.insert, e.update, e.delete, e.params)

**Advanced Schema:**

- [examples/advanced-schema.md](examples/advanced-schema.md) - Access policies, backlinks, abstract types, polymorphism, triggers

---

<philosophy>

## Philosophy

Gel is a graph-relational database. It combines the relational model (tables, constraints, ACID) with a graph model (links between objects, deep traversal). The core idea: **relationships are first-class citizens, not join tables.**

**Core principles:**

1. **Schema is the source of truth** -- Define everything in `.gel` files (or `.esdl` for legacy projects). Migrations are auto-generated by comparing your schema files against the database state.
2. **Links over foreign keys** -- Use `link` to connect types. Gel handles the underlying foreign keys. You never write JOIN -- you traverse links with dot notation.
3. **Sets everywhere** -- Every value in EdgeQL is a set. A single string is a set of one element. This is the most important mental model shift from SQL.
4. **Shapes for projection** -- SELECT returns structured, nested objects (like GraphQL responses), not flat rows. You specify the "shape" of what you want.
5. **Computed properties are views** -- Computed properties and links are not stored; they are evaluated on every query. Use them for derived data.
6. **Query builder for TypeScript** -- The generated query builder provides compile-time type safety. Prefer it over raw EdgeQL strings in TypeScript projects.

**When to use Gel:**

- Applications with complex, deeply nested relationships (social graphs, content systems, e-commerce)
- Projects that benefit from graph-style traversals without sacrificing relational integrity
- TypeScript projects that want compile-time type-safe database queries
- Teams that want automatic migration generation from schema changes

**When NOT to use:**

- Existing projects locked into raw PostgreSQL with extensive stored procedures
- Applications where SQL compatibility is the only acceptable query language (Gel 6+ has native SQL support, but EdgeQL is the primary interface)
- Environments that cannot run the Gel server process

</philosophy>

---

<patterns>

## Core Patterns

### Pattern 1: Client Setup

Create a client with `createClient()`. Connection details are auto-discovered from `gel.toml` (or `edgedb.toml`) or environment variables.

```typescript
import { createClient } from "gel";

const client = createClient(); // auto-discovers from project config
export { client };
```

Choose the right query method by expected cardinality: `query()` for sets, `querySingle()` for optional single, `queryRequiredSingle()` when result is guaranteed, `execute()` for side-effect-only statements. See [examples/core.md](examples/core.md) for full client setup patterns.

---

### Pattern 2: Schema Definition (SDL)

Schemas live in `dbschema/*.gel` files (or `*.esdl` for legacy projects). Use `required` for non-null, `link` for relationships, `constraint exclusive` for uniqueness, computed backlinks (`.<author[is Post]`) for reverse traversal.

```
# dbschema/default.gel
module default {
  type User {
    required name: str;
    required email: str { constraint exclusive; };
    multi posts := .<author[is Post];  # computed backlink
  }
  type Post {
    required title: str;
    required author: User;             # link, not uuid!
  }
}
```

**Key rule:** Always use `link` for relationships -- never raw `uuid` properties. See [examples/core.md](examples/core.md) for complete schema patterns with constraints, indexes, and enums.

---

### Pattern 3: EdgeQL Queries

SELECT uses shapes for projection (like GraphQL), INSERT assigns links via subqueries, UPDATE uses `+=`/`-=`/`:=` for multi link manipulation.

```edgeql
select User {
  name, email,
  posts: { title, status } filter .status = Status.published,
} filter .email = 'alice@example.com';
```

See [examples/core.md](examples/core.md) for SELECT/INSERT/UPDATE/DELETE patterns and parameterized queries.

---

### Pattern 4: Migrations

Gel compares your `.gel` files against the database and auto-generates migrations.

```bash
gel migration create    # generate migration from schema diff (interactive)
gel migrate             # apply pending migrations (idempotent)
npx @gel/generate edgeql-js  # regenerate query builder
```

Never edit the database with DDL directly -- always modify `.gel` files and use the migration workflow. See [examples/core.md](examples/core.md) for the full workflow including `gel watch --migrate` for prototyping.

---

### Pattern 5: Transactions

Pass `tx` (not `client`) to ALL operations inside `client.transaction()`. Using `client` inside the callback runs queries outside the transaction.

```typescript
await client.transaction(async (tx) => {
  await tx.execute(`update Account ...`); // tx, not client!
});
```

See [examples/core.md](examples/core.md) for transaction patterns and [examples/query-builder.md](examples/query-builder.md) for query builder transactions.

</patterns>

---

<red_flags>

## RED FLAGS

**High Priority Issues:**

- Using `client` instead of `tx` inside `client.transaction()` -- queries run outside the transaction and cannot be rolled back
- Forgetting to run `npx @gel/generate edgeql-js` after `gel migrate` -- query builder types are stale and TypeScript won't catch schema mismatches
- Using raw `uuid` properties instead of `link` -- defeats Gel's graph traversal and referential integrity
- Using `datetime_current()` in schema-defined computed properties -- volatile functions are forbidden in schema computeds; use `datetime_of_statement()` or `datetime_of_transaction()`

**Medium Priority Issues:**

- Bare `{}` for empty sets -- EdgeQL requires explicit type cast (`<str>{}`, `<array<int64>>[]`) because the type cannot be inferred from an empty literal
- Using `:=` when you mean `+=` on multi links in UPDATE -- `:=` replaces the entire set, `+=` adds to it, `-=` removes from it
- Not specifying `filter` on UPDATE/DELETE -- without a filter, the operation applies to ALL objects of that type
- Editing the database with DDL directly instead of through `.gel` files + migrations -- causes schema drift between files and database

**Common Mistakes:**

- Expecting element-wise behavior from set operations -- `{1, 2} + {10, 20}` produces `{11, 21, 12, 22}` (Cartesian product), not `{11, 22}`
- Forgetting that `select` on a single link returns an object (not an ID) -- you do not need to JOIN; just traverse with `.`
- Using `select count(MyType)` and expecting `querySingle` to work -- `count()` always returns exactly one value, so use `queryRequiredSingle`
- Defining a computed backlink but forgetting the type filter -- `.<author` without `[is Post]` returns all types that have an `author` link

**Gotchas & Edge Cases:**

- Computed properties are not stored -- they are re-evaluated on every query, which can be expensive for complex expressions
- `required` on a multi link means "at least one" -- an empty set violates the constraint, which can be surprising
- String concatenation uses `++` not `+` -- the `+` operator is for arithmetic only
- `LIMIT 1` does NOT make a query return a singleton for cardinality purposes -- use `filter .id = <uuid>$id` (exclusive constraint) for the query builder to infer singleton cardinality
- Multi links are unordered sets -- if you need ordering, add an `order by` in your query or use an intermediate type with an `order` property
- Backlinks (`.<link_name`) default to `multi` cardinality -- use `single` keyword explicitly if you know the relationship is one-to-one
- EdgeDB branches (v5+) are separate database copies, not lightweight references -- branching a large database takes time and disk space proportional to the data size

</red_flags>

---

<critical_reminders>

## CRITICAL REMINDERS

> **All code must follow project conventions in CLAUDE.md** (kebab-case, named exports, import ordering, `import type`, named constants)

**(You MUST run `npx @gel/generate edgeql-js` after every `gel migrate` -- the generated query builder is based on the database schema and becomes stale after migrations)**

**(You MUST cast empty sets explicitly (`<str>{}`, `<int64>{}`) -- bare `{}` is a syntax error because EdgeQL is strongly typed and cannot infer the type of an empty set)**

**(You MUST understand that all EdgeQL values are sets -- operations on multi-valued expressions produce Cartesian products, not element-wise results)**

**(You MUST pass the transaction object `tx` (not `client`) to ALL query `.run()` calls inside `client.transaction()` -- using `client` inside a transaction runs queries outside the transaction)**

**(You MUST NOT use volatile functions like `datetime_current()` in schema-defined computed properties -- use `datetime_of_transaction()` or `datetime_of_statement()` instead)**

**Failure to follow these rules will cause stale types, silent data bugs, or transaction isolation failures.**

</critical_reminders>