graphql

GraphQL

Purpose

Design a GraphQL schema that models the domain rather than the database, and operate it without letting a single query take down the service.

When to Use

• Designing a GraphQL schema from scratch.
• Diagnosing slow queries or N+1 resolver behavior.
• Protecting a public GraphQL endpoint from expensive queries.
• Splitting a schema across services with federation.

Capabilities

• Schema design: types, interfaces, unions, connections, nullability.
• Resolver architecture and DataLoader batching.
• Query cost analysis, depth limiting, and persisted queries.
• Error handling that distinguishes partial failures from total ones.
• Federation and schema composition.

Inputs

• The domain model and the client's actual query patterns.
• The data sources behind each field.
• Whether the endpoint is public (untrusted queries) or internal.

Outputs

• A schema with deliberate nullability and stable field names.
• Resolvers that batch, with no N+1 on any documented query.
• Cost limits and a persisted-query allow-list for public endpoints.

Workflow

1. Design for the client, not the tables — The schema is a product surface. If it mirrors your database, you have built a slower REST API with worse caching.
2. Get nullability right early — A nullable field is a permanent client burden; a non-null field that later fails takes down the whole parent object. Non-null for genuine invariants only.
3. Batch every relation — Every resolver that fetches by id gets a DataLoader. Without one, orders { customer { name } } issues one query per order.
4. Bound the cost — Depth limit, complexity limit, and pagination caps. Then persisted queries for first-party clients.
5. Model errors explicitly — Expected failures (validation, not found) belong in the schema as union results; unexpected failures go to errors.

Best Practices

• Never expose an unbounded list field. Use the Connection pattern with a first/after cap.
• DataLoader instances are per-request. A shared loader is a cache-poisoning bug across users.
• Changing a field from nullable to non-null is a breaking change for clients that handle null; the reverse is breaking for clients that do not. Get it right before launch.
• Do not version a GraphQL schema. Add fields, deprecate old ones with @deprecated(reason:), and remove them once usage reaches zero.
• Instrument per-field resolver latency. The slow field is never the one you would guess.
• Introspection on a public production endpoint is a reconnaissance gift. Disable it or restrict it.