ums

ADR-0055: GraphQL/REST Hybrid API Pattern

Field	Value
Status	Accepted
Date	2026-05-21
Context	UMS Web App — API Communication Strategy
Deciders	Architecture Team

Problem

The UMS system needs to support both flexible data querying (with nested relationships, filtering, and field selection) and clear transactional command semantics. Using a single API pattern for both leads to either over-fetching (REST) or unclear mutation semantics (GraphQL).

Decision

Adopt a GraphQL for Queries, REST for Commands hybrid pattern:

GraphQL (HotChocolate): All read operations (queries). Clients request exactly the fields they need, with nested relationships in a single round-trip.
REST Minimal APIs: All write operations (commands/transactions). Clear HTTP semantics (POST, PUT, DELETE) with explicit status codes and idempotency guarantees.

Client-Side Implementation

Frontend (React)
├── GraphQL Client (graphql-request v7)
│   ├── All queries use absolute URL: `${window.location.origin}/graphql`
│   ├── Typed queries generated from schema
│   └── Cached via TanStack Query
│
└── REST Client (Axios via httpClient.ts)
    ├── All mutations (POST, PUT, DELETE)
    ├── CSRF token injection for state-changing requests
    ├── Dev headers (X-User-Id, X-Language, X-Tenant-Id)
    └── Error normalization via interceptors

Rationale

Query flexibility: GraphQL allows the frontend to request exactly what each screen needs without over-fetching or N+1 requests.
Clear transaction semantics: REST provides well-understood HTTP status codes, idempotency keys, and retry semantics for mutations.
CSRF protection: REST endpoints are naturally protected via CSRF tokens; GraphQL subscriptions/queries are read-only and CSRF-safe.
TanStack Query integration: GraphQL queries cache naturally with TanStack Query’s query key system.

Consequences

Positive:

Reduced payload sizes (field selection)
Single round-trip for nested data
Clear separation of read vs write concerns
Natural CSRF protection boundary

Negative:

Two API clients to maintain
Developers must know which pattern to use
Vite proxy configuration needed for both /api and /graphql

Implementation

src/infrastructure/http/httpClient.ts — Axios instance for REST commands
src/infrastructure/http/graphqlClient.ts — GraphQL client for queries
src/infrastructure/http/csrf.ts — CSRF token management
vite.config.ts — Proxy /api and /graphql to backend

Alternatives Considered

Alternative 1: Separate API Tiers (Query Tier + Command Tier)

Split queries and commands into two independently deployed API services — a dedicated GraphQL query service and a dedicated REST command service.

Rejected because:

UMS is a modular monolith. Splitting into deployment tiers before extraction criteria are met violates ADR-0054 (Shell Library Isolation) and the modular monolith evolution playbook.
CQRS separates read and write models, not deployment units. The existing separation is already enforced at three levels: protocol (GraphQL vs REST), code (distinct handlers, distinct clients), and routing (/graphql vs /api/v1/...).
Operational cost doubles: two Dockerfiles, two health checks, two scaling policies, two sets of connection pools — with no measurable benefit at current load.

When this decision should be revisited:

This alternative becomes valid when any of the following conditions are met:

Trigger	Explanation
Read throughput consistently 10x write throughput	Independent horizontal scaling of the query tier becomes justified
Separate teams own query vs command surfaces	Conway’s Law makes the split natural, not forced
Migration toward microservices is initiated	Tier separation is a prerequisite step
Incompatible technology requirements emerge	e.g., query tier needs a different runtime or caching strategy

SaaS-specific consideration — tenant load isolation:

In a multi-tenant SaaS context, heavy GraphQL queries from a large tenant could impact command latency (login, provisioning) if both share the same process. This risk is mitigated in the current architecture by:

GraphQL query complexity limits enforced at the HotChocolate schema level.
Differentiated timeouts per operation type.
Per-tenant rate limiting at the API Gateway layer (see TE-07: YARP API Gateway).

Tier separation remains the correct escalation path if these controls prove insufficient at scale.

Alternative 2: GraphQL for Both Queries and Mutations

Use GraphQL exclusively — queries and mutations — eliminating the REST layer.