Link Search Menu Expand Document

Review of GraphQL for the Marxan cloud platform

Some key points from my deep dive into the GraphQL ecosystem.

The TL;DR

  • JSON:API and GraphQL should both be a good fit for the project
  • frontend devs should be able to reuse some existing components/strategies if using JSON:API, and have experience with JSON:API
  • the API framework we are using for Marxan (NestJS, also used in Wildlife Insights and Mars MGIS) natively supports GraphQL and we have experience and reusable code for GraphQL, from Wildlife Insights
  • JSON:API would need to be integrated into NestJS via a TypeScript implementation of the JSON:API specification (we may need to do some research via proofs of concept)
  • (we also used GraphQL via Apollo for the Wildlife Insights frontend, so we may be able to reuse code or at least strategies from this)
  • JSON:API relies on familiar HTTP and REST concepts
  • GraphQL departs in several ways from HTTP and REST concepts
  • GraphQL allows to focus on strict typing and on a schema-first approach
  • (optionally, GraphQL schema and types can integrate with TypeScript type definitions)
  • a GraphQL schema should allow us to create a mock API quickly, via graphql-faker
  • Both JSON:API and GraphQL in practice require client and server tooling

My recommendation

I recommend that we use GraphQL. Despite the likely initial higher ramp-up for frontend development, we should be able to reap benefits in terms of development ergonomics over the timespan of the project.

Bias disclaimer: although I did my homework reading the JSON:API specification, I have no production experience with JSON:API and I was very happy with GraphQL in Wildlife Insights (and in some pre-Vizz Gatsby projects).

I expect that handling JSON:API with NestJS through one of the TypeScript JSON:API implementations should be reasonably doable, but here be dragons - we won’t be able to rely on an existing NestJS plugin.

A deeper dive

First, a quick example. We will create a new Git repository/project on the SourceHut platform via its GraphQL API.

$ cat example.json
mutation createRepository($name: String!, $description: String) {
  createRepository(name: $name, visibility: PRIVATE, description: $description) {
    id
    updated
    owner {
      canonicalName
    }
  }
}

$ curl -X POST -H 'Authorization: Bearer <token>' \
    -H 'Content-Type: application/graphql'
    -d @example.json
    'https://git.sr.ht/query?variables={"name": "marxan-cloud-ftw", "description": "Components for the MarxanCloud platform"}' | jq
{
  "data": {
    "createRepository": {
      "id": 12345,
      "updated": "2020-12-09T10:07:00Z",
      "owner": {
        "canonicalName": "~hotzevzl"
      }
    }
  }
}

Although this is a very simple example, we can nevertheless observe some key traits of a typical GraphQL API which are relevant for a review of key differences from REST-ish APIs.

  • In simplest use cases, GraphQL is “just JSON over HTTP”

In practice, client and server libraries are needed to do most of the heavy lifting (similarly to client and server implementations of the JSON:API spec), but in essence the interface between GraphQL clients and servers is mostly a pipeline of JSON documents (requests and response) over HTTP: for very simple interactions, even plain requests via cURL will work fine.

  • Requests typically hit a single “endpoint”

In this example, the endpoint is https://git.sr.ht/query.

This acts effectively as an “entry point” to the entire GraphQL API, through which all the client request are channeled.

Meaningful, consistent and elegant URI design, which is typically a core part of a REST API, does not really apply to GraphQL. All the naming semantics are conveyed through the GraphQL payload (above shown as raw GraphQL query, but typically encapsulated in a JSON file which contains both query and variables).

In a REST API, a request similar to the example above, to create a new repository, would be done via something like POST /api/v1/repositories, with a JSON payload describing the resource that we are trying to create.

  • HTTP is used as a “dumb” transport channel

All requests use a single HTTP verb (typically, POST: GET may be used to query data only, not to mutate data), besides using a single “endpoint”.

This is a major departure from the consistent use of key HTTP verbs in REST APIs.

  • GraphQL is kind of RPC

In their simplest form, GraphQL queries do indeed feel declarative; for example:

query {
  me {
    canonicalName
  }
}

However, they can be augmented with parameters which may describe server-side operations of arbitrary complexity; such as plain filters in the example below (repositories(filter: { count: 2 }):

query {
  me {
    canonicalName
    repositories(filter: { count: 2 }) {
      cursor
      results {
        id, name, updated
      }
    }
  }
}

Or stateful cursors for paginated result sets such as in the example below:

query {
  me {
    canonicalName
    repositories(cursor: "<long string with cursor>") {
      cursor
      results {
        id, name, updated
      }
    }
  }
}

In my view, GraphQL allows to hit a sweet spot between the meaningful manipulation of representations of resources as typically done in REST APIs, and the potential expressiveness and flexibility of RPC interfaces, yet without forcing to work within the complexity of more elaborate RPC systems (for those still having SOAP or XML-RPC nightmares from the 2000s).

  • GraphQL enforces strongly typed schemas

Although it is indeed possible to create GraphQL schemas that only use basic scalar types (such as String, Int, Float, Boolean), in practice this would hardly be useful: most often, expressive types are created and composed.

Union types, interfaces, enums are commonly used to curate expressive API graphs.

“Making impossible states impossible” through strong typing is indeed a major advantage of the GraphQL query language.

In the example above, the createRepository() mutation accepts one required string parameter (name) and one optional string parameter (description), and makes use of an enum type for the visibility parameter.

Query payloads that don’t validate against the schema are outright rejected (and GraphQL tooling can help to avoid even creating invalid payloads, especially if coupled with TypeScript typing).

  • Caching is… different

A major point for GraphQL skepticism, but likely in practice not all that relevant in projects like the Marxan cloud platform.

GraphQL queries (as opposed to mutations and subscriptions) could indeed be cached by using the GET-verb endpoint, companion to the POST one normally used (with some caveats).

In practice, for authenticated GraphQL APIs where query results may be different for each authenticated user, even this may not be a major issue, especially when combined with application-level caching strategies, which are available at different levels of granularity (from whole-request down to individual fields).

Several common caching strategies are implemented in major GraphQL client and server implementations, making it possible to rely on existing best practices and keeping development complexity under control.

Moreover, expensive queries that need to be exposed to non-authenticated users could also be made available over ad-hoc REST endpoints, thus allowing to enjoy the benefits of HTTP-layer gateway caching.

Some more points to consider

More traits of GraphQL API practices cannot be gleaned from the initial example above. I am briefly outlining here the ones that are, in my opinion, most relevant to the Marxan cloud platform scenario.

  • Authentication is often kept as a REST subsystem

This seems to be common practice, and allows to use different established authentication strategies (e.g. password authentication, 2FA, OAuth2, etc.).

Once a token or key of some sort is obtained, this can be used in HTTP headers as commonly done in Bearer type authorization strategies or equivalent.

  • Introspection augments the type system

In practice, this is often used to provide functionality similar to what OpenAPI interfaces do; in the case of GraphQL, documentation of the whole API graph is consistently done via introspection, so the GraphQL schema serves both as parsing, validation and query processing layer as well as documentation layer (the same may be true in some OpenAPI server implementations, in practice).

Documentation on how to use a GraphQL API is typically still needed like with most non-trivial APIs, although the type system combined with introspection and meaningful naming of queries, mutations and parameters provides a convenient core documentation layer.

  • Authorization

Fine-grained access control is pretty much equally doable via REST and GraphQL.

For example, only allowing users with a specific role to set or modify specific fields, or to set them to specific values under specific circumstances, or to query specific fields.

GraphQL may make some access control rules more expressive (for example, by using access decorators in the schema definition) and easier to understand at a glance, although I would argue this is mostly a tooling concern as well as a matter of preference.

  • N+1 problem

GraphQL resolvers by default typically traverse a query graph node by node, depth-first and may process each node in isolation, which could easily lead to very inefficient database queries (defaults are actually up to server implementations - this would be a worst-case scenario).

GraphQL server implementations may provide, either as core functionality or through a library/module, lazy loading/async/batching functionality which essentially traverses the whole query graph (or parts of it), compiling a query plan and eventually executing this to load data from database through the least possible number of queries (simplifying things a bit here).

This pattern (“DataLoader”) may limit the ability to compute/estimate cost of individual resolvers, so in case of very complex query graphs, things will still need to be optimised for a sane balance between query performance and the ability to constrain queries whose computational cost is considered unacceptable for the specific context.

  • API Versioning

A common strategy seems to be evolutive versioning, or basically deprecating affected fields (the GraphQL schema description language provides a @Deprecated directive for this) and introducing their eventual replacements, while keeping the deprecated ones in place until it is possible/reasonable to break existing API contracts.

  • Errors

Besides sidestepping the semantics of HTTP verbs besides GET and POST, GraphQL also handles errors differently than via HTTP status codes. In practice, client and server implementations take care of error handling similarly to how HTTP status codes would be handled by REST frameworks. Additionally, GraphQL error handling can be granular, for example surfacing different kinds of errors as they apply to different parts of a query.

  • Federation

GraphQL APIs for distinct microservices can be federated at the API layer itself (e.g. via Apollo Federation), essentially providing a GraphQL-focused gateway that allows clients to see a single GraphQL API even though parts of its functionality may be provided by independent microservices.

We didn’t use nor needed this functionality in Wildlife Insights, but as we are likely going to rely on a small set of microservices for data processing operations, this could be a useful feature for the Marxan cloud platform. We could of course equally do some ad-hoc lightweight proxying of REST requests in the API.