Go GraphQL

GraphQL moves query planning to the client. That is the feature and the danger: one innocuous query can become ten thousand database round trips, and one over-permissive field can leak another tenant's data. Both are solved at the server, not in the schema review.

Detailed reference material, loaded on demand:

• references/errors-auth-testing.md — error presenter, field authorization directive, client-driven tests with query-count asserts.

Read a reference file only when the summary below is not enough.

1. Schema-First with gqlgen

The .graphql schema is the source of truth. gqlgen generates models, resolver stubs, and the execution layer from it.

go get -tool github.com/99designs/gqlgen
go tool gqlgen init      # once
go tool gqlgen generate  # after every schema change

# gqlgen.yml — bind generated types to your own models
models:
  User:
    model: github.com/myorg/app/internal/domain.User
  ID:
    model:
      - github.com/99designs/gqlgen/graphql.ID
      - github.com/99designs/gqlgen/graphql.Int64

Bind domain types explicitly. Left to itself gqlgen generates a parallel set of anaemic structs, and every resolver becomes a mapping function.

Commit generated code, and fail CI when it is stale:

go tool gqlgen generate && git diff --exit-code

Never edit generated.go or models_gen.go. resolver.go and the *.resolvers.go files are yours.

2. Resolvers Stay Thin

A resolver translates a GraphQL request into a service call. It contains no business logic and no SQL.

func (r *queryResolver) User(ctx context.Context, id string) (*domain.User, error) {
    u, err := r.users.Find(ctx, id)
    if errors.Is(err, domain.ErrNotFound) {
        return nil, nil // nullable field: absent, not an error
    }
    if err != nil {
        return nil, fmt.Errorf("find user %s: %w", id, err)
    }
    return u, nil
}

Inject dependencies through the Resolver struct, never through package globals:

type Resolver struct {
    users  UserService
    orders OrderService
    loader *Loaders
}

Always propagate ctx. It carries the request deadline, the authenticated principal, and the per-request dataloaders.

3. The N+1 Problem — the one that matters

A field resolver on a list type runs once per element.

// ❌ 1 query for the orders, then N queries for the users
func (r *orderResolver) Customer(ctx context.Context, obj *domain.Order) (*domain.User, error) {
    return r.users.Find(ctx, obj.CustomerID)
}

Batch with a dataloader. It collects the keys requested within a short window and issues one query.

import "github.com/vikstrous/dataloadgen"

type Loaders struct {
    UserByID *dataloadgen.Loader[string, *domain.User]
}

func NewLoaders(s UserService) *Loaders {
    return &Loaders{
        UserByID: dataloadgen.NewLoader(func(ctx context.Context, ids []string) ([]*domain.User, []error) {
            return s.FindMany(ctx, ids) // ONE query for all ids
        }, dataloadgen.WithWait(time.Millisecond)),
    }
}

// ✅ 1 query for the orders, 1 for all customers
func (r *orderResolver) Customer(ctx context.Context, obj *domain.Order) (*domain.User, error) {
    return loadersFrom(ctx).UserByID.Load(ctx, obj.CustomerID)
}

Loaders are per request, installed by middleware. A process-wide loader caches across users and leaks data between them.

func withLoaders(svc UserService, next http.Handler) http.Handler {
    return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
        ctx := context.WithValue(r.Context(), loadersKey{}, NewLoaders(svc))
        next.ServeHTTP(w, r.WithContext(ctx))
    })
}

The batch function must return results in the order of the keys it was given, with a nil entry and an error per missing key. Returning a shorter slice silently misaligns every result.

4. Bound Every Query

A public GraphQL endpoint without limits is a denial-of-service endpoint.

srv := handler.New(generated.NewExecutableSchema(cfg))
srv.AddTransport(transport.POST{})
srv.SetQueryCache(lru.New[*ast.QueryDocument](1000))
srv.Use(extension.FixedComplexityLimit(300))
srv.Use(extension.AutomaticPersistedQuery{Cache: lru.New[string](100)})

• Complexity limit — assign a cost per field, higher for list fields with a large first. Start at a number your slowest legitimate query fits under, then measure.
• Depth — recursive types (user { orders { customer { orders ... } } }) must be bounded. gqlgen has no built-in depth limit; enforce it in an operation middleware.
• Pagination is mandatory on every list field. A field returning an unbounded list is a schema bug.
• Introspection is only enabled if you install extension.Introspection. Do not install it in production, or gate it behind an authenticated role.
• Persisted queries let a public client send a hash instead of a document, so the server executes only queries you shipped.

Set srv.AroundOperations to enforce a per-operation timeout, and always run behind an http.Server with ReadTimeout and WriteTimeout set.

5. Errors

GraphQL returns 200 with an errors array. Never leak internals into it.

Install srv.SetErrorPresenter: map known domain errors (validation, not found) to their message plus a stable extensions.code, and log everything else server-side while returning internal server error with code INTERNAL. Example: references/errors-auth-testing.md.

Use srv.SetRecoverFunc to convert a resolver panic into an error instead of killing the connection, and log it with the stack.

Remember the nullability rule: an error on a non-null field nulls out its nearest nullable ancestor. Make a field non-null only when it can never legitimately be absent.

6. Authorization Belongs on the Field

Object-level checks are not enough — a client can reach an object through several paths.

Declare a @hasRole (or similar) directive on sensitive fields and implement it as a check against the principal in the context; return access denied without detail. Example: references/errors-auth-testing.md.

Authenticate in HTTP middleware, before the GraphQL handler. Authorize in the directive or the resolver, using the principal from the context — never from a query argument.

7. Testing

Drive the real schema through gqlgen's test client and assert on the decoded response. For every resolver behind a dataloader, also assert the query count — that is the only way an N+1 regression fails a build rather than a dashboard. Examples: references/errors-auth-testing.md.

Verification Checklist

1. Schema is the source of truth; generated files are committed and CI-checked
2. Generated types bind to domain models via gqlgen.yml
3. Resolvers contain no business logic and always propagate ctx
4. Every list-field resolver that fetches by ID goes through a dataloader
5. Dataloaders are constructed per request, never shared across requests
6. Batch functions return one result per key, in key order
7. A complexity limit and a depth bound are configured and tested
8. Every list field is paginated
9. Introspection is disabled or role-gated in production
10. An error presenter strips internal errors; a recover func is installed
11. Authorization is enforced per field, from the context principal
12. A test asserts the query count for at least one batched field