GraphQL or REST: how to decide for a new API

What it is

GraphQL (cited introduction) is a query language plus a server-side runtime that executes queries against a type system you define; the client sends a query shaped like the data it wants and receives exactly that data in one request, and the API evolves by adding fields and marking old ones @deprecated rather than by versioning. The GraphQL-over-HTTP guidance (cited) describes the operational shape: a single endpoint, typically /graphql; POST for queries and mutations, optionally GET for queries; and a 2xx status even when the response contains errors, because HTTP has no status code for partial success. A REST-style API instead exposes resources at URLs, uses methods and status codes for semantics, and relies on HTTP machinery for caching and errors.

Why it matters

The choice is architectural and hard to reverse. It changes how caching works, how errors are reported, how rate limits are computed and what tooling clients need.

How to apply

• Choose GraphQL when many different clients (web, mobile, partner apps) need different slices of a connected data graph, when over-fetching or request fan-out is a measured problem, and when you control the clients enough to ship a GraphQL client library.
• Choose REST when the API is resource-oriented and public, when HTTP caching by URL matters, when integrators are scripts, webhooks and agents that speak plain HTTP, or when operations tooling is built around status codes and paths.
• If choosing GraphQL, plan the controls the GraphQL security guidance (cited) describes: depth limits, limits on breadth and aliases, and query complexity analysis with per-client budgets; the specification itself does not define these. Plan resolver batching for the N+1 problem.
• If choosing REST, plan field selection, filtering and compound endpoints so clients are not forced into many round trips.
• Consider both: REST for the public surface and integrations, GraphQL as a backend-for-frontend for your own user interfaces.

Pitfalls

Adopting GraphQL to avoid API design; the schema needs the same care as resource design. Exposing GraphQL to untrusted callers without cost limits. Growing a REST API's nested-include parameters until it is an unconstrained query language without GraphQL's validation. Expecting CDN caching from GraphQL over POST; GET queries can be cached but hit URL length limits.